| Require: Data Samples D, Network f |
| 1: for (x,y) ~ D do | ▷ Cross-Fitting |
| 2: t' ← g(x) | ▷ PGD Attack |
| 3: (x1,y1,t1'), (x2,y2,t2') ~ Split(x,y,t') | |
| 4: L_a ← L_Delphi(x1,y1,t1';f) | ▷ Mild Assumption |
| 5: (xt2,yt2,t2) ← Select(y2 ≠ f(x2+t2')) ▷ Worst | |
| 6: j* ← arg maxj f_j(xt2+t2) | ▷ j: Class Index |
| 7: τ ← 1/fj*(xt2+t2)-1 | ▷ Balancing Ratio |
| 8: L_b ← τL_CE(f(xt2+t2),yt2) + L_CE(f(xt2),yt2) | |
| 9: L_ADML ← L_a + L_b | ▷ ADML Loss |
| 10: wf ← wf - α ∂/∂w_f L_ADML ▷ Weight Update (α: lr) | |
| 11: end for | |
",
+ "bbox": [
+ 501,
+ 107,
+ 890,
+ 292
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "nitude of our causal parameter also $|\\hat{\\theta}|$ becomes large due to $|\\hat{\\theta}| \\propto |1 / \\mathbb{E}_{t' | x}[p(y_a | x, t')]|$ .",
+ "bbox": [
+ 500,
+ 301,
+ 890,
+ 335
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Bringing such factors at once, $\\hat{\\theta}$ represents a weighted measurement of attacked confidence and their input gradients. Comprehensively, we can revisit that the network predictions of worst examples are easily flipped due to the following adversarial vulnerability: (a) ambiguous confidence around classification boundaries, or (b) high gradient magnitude amplifying the leverage of the perturbations. To improve the adversarial robustness of DNNs, it is essential to minimize the negative effects of causal parameters, which are combinatorial outcomes of the gradient and confidence.",
+ "bbox": [
+ 496,
+ 337,
+ 890,
+ 488
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.4. Mitigating Adversarial Vulnerability",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 500,
+ 816,
+ 517
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "By deploying ADML, we propose a way of estimating causal parameter representing the degree of adversarial vulnerability that disturbs to predict the target classes. Then, our final goal is essentially to lessen its direct causal effect from adversarial perturbations in order to achieve robust networks. In detailed, alleviating their causal effect derived from $\\hat{\\theta}$ is the process of comprehensive reconstruction to focus more on vulnerable samples as we reflect their attacked confidence and gradients effects altogether.",
+ "bbox": [
+ 496,
+ 523,
+ 890,
+ 660
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Accordingly, the very first way is naively reducing the magnitude of $\\hat{\\theta}$ to suppress adversarial vulnerability damaging the robustness. However, calculating $\\hat{\\theta}$ and minimizing its magnitude at every iteration is computationally striking because input gradient has huge dimension of $\\mathbb{R}^{dhwc}$ and getting its gradient inevitably needs to compute second-order gradient with its tremendous dimension. We instead approximate the partial derivative $\\frac{\\partial}{\\partial t}\\mathbb{E}[y|\\mathrm{do}(\\mathcal{T} = t)]$ and minimize its magnitude, which can be written as:",
+ "bbox": [
+ 496,
+ 662,
+ 890,
+ 797
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\min _ {f} | \\hat {\\theta} | \\approx \\left| \\frac {\\mathbb {E} [ y \\mid \\operatorname {d o} (\\mathcal {T} = t) ] - \\mathbb {E} [ y \\mid \\operatorname {d o} (\\mathcal {T} = 0) ]}{t - 0} \\right|, \\tag {9}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 522,
+ 809,
+ 890,
+ 844
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where network parameters of DNNs $f$ are only dependent on the numerator, thus we engross the numerator only. Lastly, we redesign $\\mathbb{E}[y|\\mathrm{do}(\\mathcal{T} = t)]$ into the form of loss",
+ "bbox": [
+ 496,
+ 854,
+ 890,
+ 901
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "4503",
+ "bbox": [
+ 480,
+ 944,
+ 514,
+ 955
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "table",
+ "img_path": "images/8ad57b2acbcb8d03d54c261188682ca29695deacf3170f4f81bfd403fc0e4eb2.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Require: Data Samples D, Network f |
| 1: for (x,y) ~ D do | ▷ Cross-Fitting |
| 2: t' ← g(x) | ▷ PGD Attack |
| 3: (x1,y1,t1'), (x2,y2,t2') ~ Split(x,y,t') | |
| 4: L_a ← L_Delphi(x1,y1,t1';f) | ▷ Mild Assumption |
| 5: (xt2,yt2,t2) ← Select(y2 ≠ f(x2+t2')) ▷ Worst | |
| 6: j* ← arg maxj f_j(xt2+t2) | ▷ j: Class Index |
| 7: τ ← 1/fj*(xt2+t2)-1 | ▷ Balancing Ratio |
| 8: L_b ← τL_CE(f(xt2+t2),yt2) + L_CE(f(xt2),yt2) | |
| 9: L_ADML ← L_a + L_b | ▷ ADML Loss |
| 10: wf ← wf - α ∂/∂w_f L_ADML ▷ Weight Update (α: lr) | |
| 11: end for | |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.5,
+ 0.303,
+ 0.892,
+ 0.336
+ ],
+ "angle": 0,
+ "content": "nitude of our causal parameter also \\( |\\hat{\\theta}| \\) becomes large due to \\( |\\hat{\\theta}| \\propto |1 / \\mathbb{E}_{t' | x}[p(y_a | x, t')]| \\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.338,
+ 0.892,
+ 0.489
+ ],
+ "angle": 0,
+ "content": "Bringing such factors at once, \\(\\hat{\\theta}\\) represents a weighted measurement of attacked confidence and their input gradients. Comprehensively, we can revisit that the network predictions of worst examples are easily flipped due to the following adversarial vulnerability: (a) ambiguous confidence around classification boundaries, or (b) high gradient magnitude amplifying the leverage of the perturbations. To improve the adversarial robustness of DNNs, it is essential to minimize the negative effects of causal parameters, which are combinatorial outcomes of the gradient and confidence."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.501,
+ 0.818,
+ 0.518
+ ],
+ "angle": 0,
+ "content": "3.4. Mitigating Adversarial Vulnerability"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.525,
+ 0.892,
+ 0.661
+ ],
+ "angle": 0,
+ "content": "By deploying ADML, we propose a way of estimating causal parameter representing the degree of adversarial vulnerability that disturbs to predict the target classes. Then, our final goal is essentially to lessen its direct causal effect from adversarial perturbations in order to achieve robust networks. In detailed, alleviating their causal effect derived from \\(\\hat{\\theta}\\) is the process of comprehensive reconstruction to focus more on vulnerable samples as we reflect their attacked confidence and gradients effects altogether."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.663,
+ 0.892,
+ 0.798
+ ],
+ "angle": 0,
+ "content": "Accordingly, the very first way is naively reducing the magnitude of \\(\\hat{\\theta}\\) to suppress adversarial vulnerability damaging the robustness. However, calculating \\(\\hat{\\theta}\\) and minimizing its magnitude at every iteration is computationally striking because input gradient has huge dimension of \\(\\mathbb{R}^{dhwc}\\) and getting its gradient inevitably needs to compute second-order gradient with its tremendous dimension. We instead approximate the partial derivative \\(\\frac{\\partial}{\\partial t}\\mathbb{E}[y|\\mathrm{do}(\\mathcal{T} = t)]\\) and minimize its magnitude, which can be written as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.523,
+ 0.81,
+ 0.892,
+ 0.845
+ ],
+ "angle": 0,
+ "content": "\\[\n\\min _ {f} | \\hat {\\theta} | \\approx \\left| \\frac {\\mathbb {E} [ y \\mid \\operatorname {d o} (\\mathcal {T} = t) ] - \\mathbb {E} [ y \\mid \\operatorname {d o} (\\mathcal {T} = 0) ]}{t - 0} \\right|, \\tag {9}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.856,
+ 0.892,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "where network parameters of DNNs \\(f\\) are only dependent on the numerator, thus we engross the numerator only. Lastly, we redesign \\(\\mathbb{E}[y|\\mathrm{do}(\\mathcal{T} = t)]\\) into the form of loss"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.482,
+ 0.945,
+ 0.516,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "4503"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.096,
+ 0.061,
+ 0.871,
+ 0.462
+ ],
+ "angle": 0,
+ "content": "| Require: Data Samples D, Network f |
| 1: for (x,y) ~ D do | ▷ Cross-Fitting |
| 2: t' ← g(x) | ▷ PGD Attack |
| 3: (x1,y1,t1'), (x2,y2,t2') ~ Split(x,y,t') | |
| 4: L_a ← L_Delphi(x1,y1,t1';f) | ▷ Mild Assumption |
| 5: (xt2,yt2,t2) ← Select(y2 ≠ f(x2+t2')) ▷ Worst | |
| 6: j* ← arg maxj f_j(xt2+t2) | ▷ j: Class Index |
| 7: τ ← 1/fj*(xt2+t2)-1 | ▷ Balancing Ratio |
| 8: L_b ← τL_CE(f(xt2+t2),yt2) + L_CE(f(xt2),yt2) | |
| 9: L_ADML ← L_a + L_b | ▷ ADML Loss |
| 10: wf ← wf - α ∂/∂w_f L_ADML ▷ Weight Update (α: lr) | |
| 11: end for | |
+
+nitude of our causal parameter also $|\hat{\theta}|$ becomes large due to $|\hat{\theta}| \propto |1 / \mathbb{E}_{t' | x}[p(y_a | x, t')]|$ .
+
+Bringing such factors at once, $\hat{\theta}$ represents a weighted measurement of attacked confidence and their input gradients. Comprehensively, we can revisit that the network predictions of worst examples are easily flipped due to the following adversarial vulnerability: (a) ambiguous confidence around classification boundaries, or (b) high gradient magnitude amplifying the leverage of the perturbations. To improve the adversarial robustness of DNNs, it is essential to minimize the negative effects of causal parameters, which are combinatorial outcomes of the gradient and confidence.
+
+# 3.4. Mitigating Adversarial Vulnerability
+
+By deploying ADML, we propose a way of estimating causal parameter representing the degree of adversarial vulnerability that disturbs to predict the target classes. Then, our final goal is essentially to lessen its direct causal effect from adversarial perturbations in order to achieve robust networks. In detailed, alleviating their causal effect derived from $\hat{\theta}$ is the process of comprehensive reconstruction to focus more on vulnerable samples as we reflect their attacked confidence and gradients effects altogether.
+
+Accordingly, the very first way is naively reducing the magnitude of $\hat{\theta}$ to suppress adversarial vulnerability damaging the robustness. However, calculating $\hat{\theta}$ and minimizing its magnitude at every iteration is computationally striking because input gradient has huge dimension of $\mathbb{R}^{dhwc}$ and getting its gradient inevitably needs to compute second-order gradient with its tremendous dimension. We instead approximate the partial derivative $\frac{\partial}{\partial t}\mathbb{E}[y|\mathrm{do}(\mathcal{T} = t)]$ and minimize its magnitude, which can be written as:
+
+$$
+\min _ {f} | \hat {\theta} | \approx \left| \frac {\mathbb {E} [ y \mid \operatorname {d o} (\mathcal {T} = t) ] - \mathbb {E} [ y \mid \operatorname {d o} (\mathcal {T} = 0) ]}{t - 0} \right|, \tag {9}
+$$
+
+where network parameters of DNNs $f$ are only dependent on the numerator, thus we engross the numerator only. Lastly, we redesign $\mathbb{E}[y|\mathrm{do}(\mathcal{T} = t)]$ into the form of loss
+
+| Require: Data Samples D, Network f |
| 1: for (x,y) ~ D do | ▷ Cross-Fitting |
| 2: t' ← g(x) | ▷ PGD Attack |
| 3: (x1,y1,t1'), (x2,y2,t2') ~ Split(x,y,t') | |
| 4: L_a ← L_Delphi(x1,y1,t1';f) | ▷ Mild Assumption |
| 5: (xt2,yt2,t2) ← Select(y2 ≠ f(x2+t2')) ▷ Worst | |
| 6: j* ← arg maxj f_j(xt2+t2) | ▷ j: Class Index |
| 7: τ ← 1/fj*(xt2+t2)-1 | ▷ Balancing Ratio |
| 8: L_b ← τL_CE(f(xt2+t2),yt2) + L_CE(f(xt2),yt2) | |
| 9: L_ADML ← L_a + L_b | ▷ ADML Loss |
| 10: wf ← wf - α ∂/∂w_f L_ADML ▷ Weight Update (α: lr) | |
| 11: end for | |
",
+ "image_path": "278a81867ac2b1a3578af499f67b14682bb887ab26d13d30433a2034dfa6bb62.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 306,
+ 239,
+ 545,
+ 266
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 239,
+ 545,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 239,
+ 545,
+ 266
+ ],
+ "type": "text",
+ "content": "nitude of our causal parameter also "
+ },
+ {
+ "bbox": [
+ 306,
+ 239,
+ 545,
+ 266
+ ],
+ "type": "inline_equation",
+ "content": "|\\hat{\\theta}|"
+ },
+ {
+ "bbox": [
+ 306,
+ 239,
+ 545,
+ 266
+ ],
+ "type": "text",
+ "content": " becomes large due to "
+ },
+ {
+ "bbox": [
+ 306,
+ 239,
+ 545,
+ 266
+ ],
+ "type": "inline_equation",
+ "content": "|\\hat{\\theta}| \\propto |1 / \\mathbb{E}_{t' | x}[p(y_a | x, t')]|"
+ },
+ {
+ "bbox": [
+ 306,
+ 239,
+ 545,
+ 266
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 304,
+ 267,
+ 545,
+ 387
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 267,
+ 545,
+ 387
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 267,
+ 545,
+ 387
+ ],
+ "type": "text",
+ "content": "Bringing such factors at once, "
+ },
+ {
+ "bbox": [
+ 304,
+ 267,
+ 545,
+ 387
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{\\theta}"
+ },
+ {
+ "bbox": [
+ 304,
+ 267,
+ 545,
+ 387
+ ],
+ "type": "text",
+ "content": " represents a weighted measurement of attacked confidence and their input gradients. Comprehensively, we can revisit that the network predictions of worst examples are easily flipped due to the following adversarial vulnerability: (a) ambiguous confidence around classification boundaries, or (b) high gradient magnitude amplifying the leverage of the perturbations. To improve the adversarial robustness of DNNs, it is essential to minimize the negative effects of causal parameters, which are combinatorial outcomes of the gradient and confidence."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 306,
+ 396,
+ 500,
+ 410
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 396,
+ 500,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 396,
+ 500,
+ 410
+ ],
+ "type": "text",
+ "content": "3.4. Mitigating Adversarial Vulnerability"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 304,
+ 415,
+ 545,
+ 523
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 415,
+ 545,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 415,
+ 545,
+ 523
+ ],
+ "type": "text",
+ "content": "By deploying ADML, we propose a way of estimating causal parameter representing the degree of adversarial vulnerability that disturbs to predict the target classes. Then, our final goal is essentially to lessen its direct causal effect from adversarial perturbations in order to achieve robust networks. In detailed, alleviating their causal effect derived from "
+ },
+ {
+ "bbox": [
+ 304,
+ 415,
+ 545,
+ 523
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{\\theta}"
+ },
+ {
+ "bbox": [
+ 304,
+ 415,
+ 545,
+ 523
+ ],
+ "type": "text",
+ "content": " is the process of comprehensive reconstruction to focus more on vulnerable samples as we reflect their attacked confidence and gradients effects altogether."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "text",
+ "content": "Accordingly, the very first way is naively reducing the magnitude of "
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{\\theta}"
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "text",
+ "content": " to suppress adversarial vulnerability damaging the robustness. However, calculating "
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{\\theta}"
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "text",
+ "content": " and minimizing its magnitude at every iteration is computationally striking because input gradient has huge dimension of "
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbb{R}^{dhwc}"
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "text",
+ "content": " and getting its gradient inevitably needs to compute second-order gradient with its tremendous dimension. We instead approximate the partial derivative "
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "inline_equation",
+ "content": "\\frac{\\partial}{\\partial t}\\mathbb{E}[y|\\mathrm{do}(\\mathcal{T} = t)]"
+ },
+ {
+ "bbox": [
+ 304,
+ 525,
+ 545,
+ 632
+ ],
+ "type": "text",
+ "content": " and minimize its magnitude, which can be written as:"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 320,
+ 641,
+ 545,
+ 669
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 320,
+ 641,
+ 545,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 320,
+ 641,
+ 545,
+ 669
+ ],
+ "type": "interline_equation",
+ "content": "\\min _ {f} | \\hat {\\theta} | \\approx \\left| \\frac {\\mathbb {E} [ y \\mid \\operatorname {d o} (\\mathcal {T} = t) ] - \\mathbb {E} [ y \\mid \\operatorname {d o} (\\mathcal {T} = 0) ]}{t - 0} \\right|, \\tag {9}",
+ "image_path": "e93e37cea9512dca8e491571de09a23fbe544064706a7be1e615e0b218a6f072.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 304,
+ 677,
+ 545,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 677,
+ 545,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 677,
+ 545,
+ 714
+ ],
+ "type": "text",
+ "content": "where network parameters of DNNs "
+ },
+ {
+ "bbox": [
+ 304,
+ 677,
+ 545,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "f"
+ },
+ {
+ "bbox": [
+ 304,
+ 677,
+ 545,
+ 714
+ ],
+ "type": "text",
+ "content": " are only dependent on the numerator, thus we engross the numerator only. Lastly, we redesign "
+ },
+ {
+ "bbox": [
+ 304,
+ 677,
+ 545,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbb{E}[y|\\mathrm{do}(\\mathcal{T} = t)]"
+ },
+ {
+ "bbox": [
+ 304,
+ 677,
+ 545,
+ 714
+ ],
+ "type": "text",
+ "content": " into the form of loss"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 315,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "text",
+ "content": "4503"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 58,
+ 48,
+ 533,
+ 365
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 58,
+ 48,
+ 533,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 58,
+ 48,
+ 533,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 58,
+ 48,
+ 533,
+ 365
+ ],
+ "type": "table",
+ "html": "| En-Es | Transcription: and always as a child I had this fantasy that somebody would come and rescue me\nMixSpeech: and always as a child I had this fantasy that somebody would come and rescue me |
| Ground Truth: y de niña siempre tenía la fantasía de que alguien vendrá a salvar me de\nMixSpeech: y de (desde) niña niño siempre tenía la esta fantasía de que alguien vendrá a salvar me (rescató) de |
| En-Fr | Transcription: and solutions create new problems which have to be solved in their turn\nMixSpeech: and solutions to create new problems which have to be solved in their turn this year |
| Ground Truth: et les solutions créent de nouveaux problèmes devant être résolus à leur tour\nMixSpeech: et les (des) solutions pour creer de nouveaux problèmes devant qui doivent être résolus à leur tour |
| En-It | Transcription: one of the last 10,000 years, and the other certainly of the last 25 years\nMixSpeech: one of the last 10,000 years and the other certainly assistancee of the last 25 years |
| Ground Truth: uno presente negli ultimi diecimila anni e l'alto certamente negli ultimi 25 anni\nMixSpeech: uno presente negli ultimi diecimila (10000) anni e l'alto certamente negli ultimi 25 anni |
| En-Pt | Transcription: has around 2,000 people descending on MIT's campus\nMixSpeech: has around 2,000 people descending on MIT's campus |
| Ground Truth: congregamerca de 2000 pessoas no campus do mit\nMixSpeech: congregam de circa de 2000 pessoas a estudar no campus do mit |
",
+ "bbox": [
+ 76,
+ 89,
+ 898,
+ 481
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Table 5. Qualitative performance of Visual Speech Recognition and Translation on AVMuST-TED. Red Strikeout Words: mistranslated words with opposite meaning, (Blue Words in parentheses): mistranslated words with similar meaning, Gray Words: the absent words.",
+ "bbox": [
+ 75,
+ 488,
+ 892,
+ 517
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "4.7. Qualitative results",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 541,
+ 254,
+ 556
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We present several examples of lip translation in Table 5 to qualitatively evaluate the translation quality of MixSpeech. The translation results are very close to the ground truth, and the semantics are consistent. We observe two types of words that differ in translation: synonyms and context-sensitive translations. Synonyms that have different spellings but the same meaning, such as salvar and rescató in Spanish, both meaning 'rescue', and diecimila and 10000 in Italian, both meaning 'ten thousand', are commonly found in translation tasks and can affect translation consistency. Additionally, there are translations that require context information, such as when the speaker refers to themselves as a child, and the translation in Spanish needs to take into account the speaker's gender to choose between niña for girl' or Niño for boy' and 'child'. The qualitative translation results of MixSpeech demonstrate its capability to achieve reliable cross-lingua lip translation. In Appendix C.2, we also provide translation results of noisy audio speech translation with visual speech translation and audio speech translation with audiovisual speech translation to highlight the importance of visual speech in speech translation.",
+ "bbox": [
+ 75,
+ 568,
+ 472,
+ 902
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "5. Conclusion",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 540,
+ 617,
+ 556
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "With the advancement of online technologies, such as online healthcare and sales, language barriers often prevent these tools from reaching and benefiting disadvantaged areas. In light of this, we focus on visual speech, a branch of the speech stream, and aim to translate visual speech from source languages to other target languages for cross-linguistic communication, specifically through lip translation. We meticulously curate the AVMuST-TED dataset, consisting of 706 hours of speech clips with professional translations from TED, to facilitate cross-linguistic research on visual speech. We also introduce MixSpeech, a cross-modality self-learning framework that utilizes mixed speech to regularize visual speech translation and achieves state-of-the-art performance in lip translation on AVMuST-TED and lip reading on LRS2, LRS3, and CMLR datasets.",
+ "bbox": [
+ 496,
+ 566,
+ 890,
+ 792
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Moreover, our work on visual speech and AVMuSTED lay a solid foundation for further research on visual speech in cross-lingual fields. There are numerous related tasks with great potential for practical applications, such as Cross-Lingual Talking Head Generation [44]. These tasks hold immense promise for breaking down language barriers and promoting communication across diverse communities.",
+ "bbox": [
+ 496,
+ 794,
+ 890,
+ 900
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "page_number",
+ "text": "15742",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Acknowledgements",
+ "text_level": 1,
+ "bbox": [
+ 78,
+ 90,
+ 243,
+ 107
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "This work was supported in part by National Key R&D Program of China under Grant No.2022ZD0162000, National Natural Science Foundation of China under Grant No.62222211, Grant No.61836002 and Grant No.62072397.",
+ "bbox": [
+ 76,
+ 114,
+ 468,
+ 189
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "References",
+ "text_level": 1,
+ "bbox": [
+ 78,
+ 203,
+ 171,
+ 218
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 1, 2, 5, 6, 12, 13",
+ "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 2, 5, 12, 13",
+ "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 6",
+ "[4] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 1",
+ "[5] Kartik Audhkhasi, Andrew Rosenberg, Abhinav Sethy, Bhuvana Ramabhadran, and Brian Kingsbury. End-to-end asr-free keyword search from speech. IEEE Journal of Selected Topics in Signal Processing, 11(8):1351-1359, 2017. 2",
+ "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 2",
+ "[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 2, 3, 4",
+ "[8] Alexandre Bérard, Olivier Pietquin, Christophe Servan, and Laurent Besacier. Listen and translate: A proof of concept for end-to-end speech-to-text translation. arXiv preprint arXiv:1612.01744, 2016. 1",
+ "[9] Peter F Brown, John Cocke, Stephen A Della Pietra, Vincent J Della Pietra, Frederick Jelinek, John Lafferty, Robert L Mercer, and Paul S Roossin. A statistical approach to machine translation. Computational linguistics, 16(2):79-85, 1990. 1",
+ "[10] Jiaao Chen, Zichao Yang, and Diyi Yang. Mixtext: Linguistically-informed interpolation of hidden space for semi-supervised text classification. arXiv preprint arXiv:2004.12239, 2020. 3",
+ "[11] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. Opensr: Open-modality speech recognition via maintaining multi-modality alignment. arXiv preprint arXiv:2306.06410, 2023. 2",
+ "[12] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 13"
+ ],
+ "bbox": [
+ 78,
+ 227,
+ 468,
+ 898
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[13] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 2, 12",
+ "[14] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. An audio-visual corpus for speech perception and automatic speech recognition. The Journal of the Acoustical Society of America, 120(5):2421-2424, 2006. 2",
+ "[15] Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672, 2022. 1",
+ "[16] Mattia A Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. Must-c: a multilingual speech translation corpus. In 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2012-2017. Association for Computational Linguistics, 2019. 5",
+ "[17] Qianqian Dong, Fengpeng Yue, Tom Ko, Mingxuan Wang, Qibing Bai, and Yu Zhang. Leveraging pseudo-labeled data to improve direct speech-to-speech translation. arXiv preprint arXiv:2205.08993, 2022. 1",
+ "[18] Qingkai Fang and Yang Feng. Neural machine translation with phrase-level universal visual representations. arXiv preprint arXiv:2203.10299, 2022. 3",
+ "[19] Qingkai Fang, Rong Ye, Lei Li, Yang Feng, and Mingxuan Wang. Stemm: Self-learning with speech-text manifold mixup for speech translation. arXiv preprint arXiv:2203.10426, 2022. 1, 2, 3",
+ "[20] Hongyu Guo, Yongyi Mao, and Richong Zhang. Mixup as locally linear out-of-manifold regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3714-3722, 2019. 3",
+ "[21] Naomi Harte and Eoin Gillen. Tcd-limit: An audio-visual corpus of continuous speech. IEEE Transactions on Multimedia, 17(5):603-615, 2015. 2",
+ "[22] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 2",
+ "[23] Wei-Ning Hsu and Bowen Shi. A single self-supervised model for many speech modalities enables zero-shot modality transfer. arXiv preprint arXiv:2207.07036, 2022.1",
+ "[24] Chenxu Hu, Qiao Tian, Tingle Li, Wang Yuping, Yuxuan Wang, and Hang Zhao. Neural dubber: Dubbing for videos according to scripts. Advances in Neural Information Processing Systems, 34:16582-16595, 2021. 1, 3",
+ "[25] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022. 2",
+ "[26] Rongjie Huang, Huadai Liu, Xize Cheng, Yi Ren, Linjun Li, Zhenhui Ye, Jinzheng He, Lichao Zhang, Jinglin Liu, Xiang"
+ ],
+ "bbox": [
+ 501,
+ 92,
+ 890,
+ 901
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "page_number",
+ "text": "15743",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "Yin, et al. Av-transpeech: Audio-visual robust speech-to-speech translation. arXiv preprint arXiv:2305.15403, 2023. 2",
+ "[27] Rongjie Huang, Yi Ren, Jinglin Liu, Chenye Cui, and Zhou Zhao. Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech. In Advances in Neural Information Processing Systems. 3",
+ "[28] Rongjie Huang, Zhou Zhao, Huadai Liu, Jinglin Liu, Chenye Cui, and Yi Ren. Prodiff: Progressive fast diffusion model for high-quality text-to-speech. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2595-2605, 2022. 2",
+ "[29] Rongjie Huang, Zhou Zhao, Jinglin Liu, Huadai Liu, Yi Ren, Lichao Zhang, and Jinzheng He. Transpeech: Speech-to-speech translation with bilateral perturbation. arXiv preprint arXiv:2205.12523, 2022. 1",
+ "[30] Taejun Kim and Juhan Nam. Temporal feedback convolutional recurrent neural networks for keyword spotting. arXiv preprint arXiv:1911.01803, 2019. 2",
+ "[31] Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 12",
+ "[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 1",
+ "[33] Ann Lee, Peng-Jen Chen, Changhan Wang, Jiatao Gu, Xutai Ma, Adam Polyak, Yossi Adi, Qing He, Yun Tang, Juan Pino, et al. Direct speech-to-speech translation with discrete units. arXiv preprint arXiv:2107.05604, 2021. 1",
+ "[34] Linjun Li, Tao Jin, Xize Cheng, Ye Wang, Wang Lin, Rongjie Huang, and Zhou Zhao. Contrastive token-wise meta-learning for unseen performer visual temporal-aligned translation. In Findings of the Association for Computational Linguistics: ACL 2023, pages 10993-11007, 2023. 2",
+ "[35] Rainer Lienhart. Reliable transition detection in videos: A survey and practitioner's guide. International journal of image and graphics, 1(03):469-486, 2001. 12",
+ "[36] Zhijie Lin, Zhou Zhao, Haoyuan Li, Jinglin Liu, Meng Zhang, Xingshan Zeng, and Xiaofei He. Simullr: Simultaneous lip reading transducer with attention-guided adaptive memory. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1359-1367, 2021. 1",
+ "[37] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 1",
+ "[38] Yuchen Liu, Hao Xiong, Zhongjun He, Jiajun Zhang, Hua Wu, Haifeng Wang, and Chengqing Zong. End-to-end speech translation with knowledge distillation. arXiv preprint arXiv:1904.08075, 2019. 1",
+ "[39] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. Lira: Learning visual speech representations from audio through self-supervision. arXiv preprint arXiv:2106.09171, 2021. 2, 12"
+ ],
+ "bbox": [
+ 78,
+ 90,
+ 468,
+ 900
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[40] Shihui Ma, Shilin Wang, and Xiang Lin. A transformer-based model for sentence-level chinese mandarin lipreading. In 2020 IEEE Fifth International Conference on Data Science in Cyberspace (DSC), pages 78-81. IEEE, 2020. 6",
+ "[41] ML Menéndez, JA Pardo, L Pardo, and MC Pardo. The jensen-shannon divergence. Journal of the Franklin Institute, 334(2):307-318, 1997. 4",
+ "[42] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. arXiv preprint arXiv:2009.01225, 2020. 2",
+ "[43] David S Moore. Uncertainty. On the shoulders of giants: New approaches to numeracy, pages 95-137, 1990. 3, 4",
+ "[44] NVIDIA. Nvidia maxine: Reinventing real-time video communications with ai. 2022. 1, 2, 8",
+ "[45] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 2, 5",
+ "[46] Matt Post. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Belgium, Brussels, Oct. 2018. Association for Computational Linguistics. 5",
+ "[47] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 2, 6",
+ "[48] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1",
+ "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3",
+ "[50] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6",
+ "[51] Max Ritter, Uwe Meier, Jie Yang, and Alex Waibel. Face translation: A multimodal translation agent. In AVSP'99-International Conference on Auditory-Visual Speech Processing, 1999. 2",
+ "[52] Andrew Rosenberg, Kartik Audhkhasi, Abhinav Sethy, Bhuvana Ramabhadran, and Michael Picheny. End-to-end speech recognition and keyword search on low-resource languages. In 2017 IEEE international conference on acoustics, speech and signal processing (icassp), pages 5280-5284. IEEE, 2017. 2",
+ "[53] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. arXiv preprint arXiv:2201.02184, 2022. 1, 2, 3, 4, 5, 6, 12, 13"
+ ],
+ "bbox": [
+ 501,
+ 92,
+ 890,
+ 900
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "page_number",
+ "text": "15744",
+ "bbox": [
+ 480,
+ 944,
+ 519,
+ 955
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[54] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 1, 13",
+ "[55] David Snyder, Guoguo Chen, and Daniel Povey. Musan: A music, speech, and noise corpus. ArXiv, abs/1510.08484, 2015. 6, 13",
+ "[56] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 6",
+ "[57] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21425–21430, 2022. 1, 2",
+ "[58] Matthias Sperber, Graham Neubig, Jan Niehues, and Alex Waibel. Neural lattice-to-sequence models for uncertain inputs. arXiv preprint arXiv:1704.00559, 2017. 1, 2",
+ "[59] Yun Tang, Juan Pino, Changhan Wang, Xutai Ma, and Dmitriy Genzel. A general multi-task learning framework to leverage text data for speech to text tasks. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6209-6213. IEEE, 2021. 3",
+ "[60] Alexander Waibel, Moritz Behr, Fevziye Irem Eyiokur, Dogucan Yaman, Tuan-Nam Nguyen, Carlos Mullov, Mehmet Arif Demirtas, Alperen Kantarcı, Stefan Constantin, and Hazim Kemal Ekenel. Face-dubbing++: Lipsynchronous, voice preserving translation of videos. arXiv preprint arXiv:2206.04523, 2022. 1, 2",
+ "[61] Chen Wang, Yuchen Liu, Boxing Chen, Jiajun Zhang, Wei Luo, Zhongqiang Huang, and Chengqing Zong. Discrete cross-modal alignment enables zero-shot speech translation. arXiv preprint arXiv:2210.09556, 2022. 1",
+ "[62] Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. arXiv preprint arXiv:2101.00390, 2021.5",
+ "[63] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 3",
+ "[64] Zehan Wang, Yang Zhao, Xize Cheng, Haifeng Huang, Jiageng Liu, Li Tang, Linjun Li, Yongqi Wang, Aoxiong Yin, Zhang Zhang, et al. Connecting multi-modal contrastive representations. arXiv preprint arXiv:2305.14381, 2023. 2",
+ "[65] Shuang Yang, Yuanhang Zhang, Dalu Feng, Mingmin Yang, Chenhao Wang, Jingyun Xiao, Keyu Long, Shiguang Shan, and Xilin Chen. Lrw-1000: A naturally-distributed large-scale benchmark for lip reading in the wild. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1–8. IEEE, 2019. 12"
+ ],
+ "bbox": [
+ 78,
+ 90,
+ 467,
+ 898
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "[66] Rong Ye, Mingxuan Wang, and Lei Li. End-to-end speech translation via cross-modal progressive training. arXiv preprint arXiv:2104.10380, 2021. 2",
+ "[67] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3",
+ "[68] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, Amirata Ghorbani, and James Zou. How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819, 2020. 3",
+ "[69] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 6",
+ "[70] Ya Zhao, Rui Xu, and Mingli Song. A cascade sequence-to-sequence model for chinese mandarin lip reading. In Proceedings of the ACM Multimedia Asia, pages 1-6. 2019. 5, 6",
+ "[71] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6917-6924, 2020. 2, 5, 6"
+ ],
+ "bbox": [
+ 501,
+ 92,
+ 890,
+ 415
+ ],
+ "page_idx": 10
+ },
+ {
+ "type": "page_number",
+ "text": "15745",
+ "bbox": [
+ 480,
+ 945,
+ 517,
+ 955
+ ],
+ "page_idx": 10
+ }
+]
\ No newline at end of file
diff --git a/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/8115edb8-8391-4f40-bf06-0cd9421711d5_model.json b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/8115edb8-8391-4f40-bf06-0cd9421711d5_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..a2936c35c83078cebb56991805b8d7999cd3396c
--- /dev/null
+++ b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/8115edb8-8391-4f40-bf06-0cd9421711d5_model.json
@@ -0,0 +1,2191 @@
+[
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.107,
+ 0.003,
+ 0.182,
+ 0.043
+ ],
+ "angle": 0,
+ "content": "CVF"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.239,
+ 0.001,
+ 0.808,
+ 0.047
+ ],
+ "angle": 0,
+ "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.102,
+ 0.13,
+ 0.87,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "MixSpeech: Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.184,
+ 0.204,
+ 0.793,
+ 0.275
+ ],
+ "angle": 0,
+ "content": "Xize Cheng*, Tao Jin*, Rongjie Huang*, Linjun Li \nWang Lin, Zehan Wang, Ye Wang, Huadai Liu, Aoxiong Yin, Zhou Zhao† \nZhejiang University \n{chengxize, jint.zju, rongjiehuang, zhaozhou}@zju.edu.cn"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.235,
+ 0.309,
+ 0.314,
+ 0.327
+ ],
+ "angle": 0,
+ "content": "Abstract"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.342,
+ 0.474,
+ 0.705
+ ],
+ "angle": 0,
+ "content": "Multi-media communications facilitate global interaction among people. However, despite researchers exploring cross-lingual translation techniques such as machine translation and audio speech translation to overcome language barriers, there is still a shortage of cross-lingual studies on visual speech. This lack of research is mainly due to the absence of datasets containing visual speech and translated text pairs. In this paper, we present AVMuST-TED, the first dataset for Audio-Visual Multilingual Speech Translation, derived from TED talks. Nonetheless, visual speech is not as distinguishable as audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this issue, we propose MixSpeech, a cross-modality self-learning framework that utilizes audio speech to regularize the training of visual speech tasks. To further minimize the cross-modality gap and its impact on knowledge transfer, we suggest adopting mixed speech, which is created by interpolating audio and visual streams, along with a curriculum learning strategy to adjust the mixing ratio as needed. MixSpeech enhances speech translation in noisy environments, improving BLEU scores for four languages on AVMuST-TED by \\(+1.4\\) to \\(+4.2\\). Moreover, it achieves state-of-the-art performance in lip reading on CMLR \\((11.1\\%)\\), LRS2 \\((25.5\\%)\\), and LRS3 \\((28.0\\%)\\)."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.734,
+ 0.21,
+ 0.75
+ ],
+ "angle": 0,
+ "content": "1. Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.759,
+ 0.47,
+ 0.866
+ ],
+ "angle": 0,
+ "content": "Multi-media techniques, including Audio-Visual Speech Recognition (AVSR) [4, 1, 2, 53], Audio-Visual Speech Translation (AVST) [8, 38, 61], and Audio-Visual Speech Generation (AVSG) [48, 32, 24], are commonly employed in various online communication scenarios, such as conferences, education, and healthcare. As a tool for ultra-remote communication, many online interactions involve multiple"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.511,
+ 0.326,
+ 0.738,
+ 0.449
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.747,
+ 0.334,
+ 0.879,
+ 0.346
+ ],
+ "angle": 0,
+ "content": "En: Hello, beautiful world."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.747,
+ 0.378,
+ 0.882,
+ 0.39
+ ],
+ "angle": 0,
+ "content": "Es: Hola, hermoso mundo."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.747,
+ 0.394,
+ 0.881,
+ 0.405
+ ],
+ "angle": 0,
+ "content": "Fr: Bonjour, beau monde."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.747,
+ 0.409,
+ 0.857,
+ 0.421
+ ],
+ "angle": 0,
+ "content": "It : Ciao, bel mondo."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.747,
+ 0.425,
+ 0.857,
+ 0.437
+ ],
+ "angle": 0,
+ "content": "Pt: Olá, belo mundo."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.747,
+ 0.44,
+ 0.783,
+ 0.448
+ ],
+ "angle": 0,
+ "content": "···"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.498,
+ 0.462,
+ 0.895,
+ 0.518
+ ],
+ "angle": 0,
+ "content": "Figure 1. Diagram of speech tasks. Audio speech and visual speech are paired parallel speech streams which can be employed for speech recognition and speech translation. However, only Lip-Translation remains unexplored."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.55,
+ 0.895,
+ 0.746
+ ],
+ "angle": 0,
+ "content": "languages, prompting the need for addressing cross-lingual challenges. Several works have attempted to tackle these challenges, including Machine Translation (MT) [9, 37, 15] for text utterance, Speech Translation (ST) [58, 19] for audio utterance, and Speech-to-Speech Translation (S2ST) [58, 19, 17, 33, 29] for simultaneous interpretation. However, research on cross-lingual visual speech is still limited, as illustrated in Figure 1. As an essential component of multi-media speech, visual speech can be combined with audio to enhance the recognition and understanding of speech content as audio-visual speech [1, 2, 54], and is the unique resource for speech content understanding in audio-disabled scenarios [36]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.75,
+ 0.896,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "Visual speech translation has never been studied, mainly for the lack of visual speech datasets with translated texts in different languages. The few remaining works [57, 60, 44] also cannot be quantitatively verified for this reason, making them unconvincing. The available visual speech corpus is often very scarce compared to audio speech owing to the high demands of visual speech for model training, which requires mostly-frontal and high-resolution videos with a sufficiently high frame rate, such that motions around the lip area are clearly captured [23]. In this paper, we propose the"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.096,
+ 0.875,
+ 0.208,
+ 0.888
+ ],
+ "angle": 0,
+ "content": "*Equal contribution."
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.097,
+ 0.888,
+ 0.222,
+ 0.9
+ ],
+ "angle": 0,
+ "content": "† Corresponding author."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.096,
+ 0.875,
+ 0.222,
+ 0.9
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.519,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "15735"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.092,
+ 0.473,
+ 0.257
+ ],
+ "angle": 0,
+ "content": "first Audio-Visual Multilingual Speech Translation dataset, AVMuST-TED. During the process of acquisition, we first screen out videos with professional translations in four different languages from TED talk which performs strict translation and review processes, and then determine the real speaker's talking head by checking whether each pair of visual speech (i.e., talking head) and audio speech matches in the manner of [1, 2]. Incidentally, this dataset can also be used for quantitative evaluation of other multi-modality translation tasks, such as cross-lingua audio-visual speech generation [51, 60]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.259,
+ 0.473,
+ 0.606
+ ],
+ "angle": 0,
+ "content": "The cascaded model comprising of a speech recognition model and a machine translation model can handle speech translation tasks but suffers from error accumulation due to model cascades and cannot process languages without text (e.g., Minnan). Our proposed end-to-end model, which can translate directly from source speech to target text, addresses the above issues. However, visual speech is less distinguishable than audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this, we introduce MixSpeech, a method that first pretrains the decoder using high-discrimination audio speech to obtain a mapping from speech phonemes to text and then generalizes this mapping to the visual speech task through cross-modality self-learning. Furthermore, since audio speech and visual speech are two distinct modalities of speech, there is a significant modality gap between them that hinders knowledge transfer. To narrow this gap and improve knowledge transfer, we propose mixed speech, which is created by interpolating audio and visual streams, rather than relying solely on audio speech. We also propose a curriculum-learning [7] based strategy to adjust the mixing ratio as the training progresses and cross-modality integration deepens."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.607,
+ 0.47,
+ 0.639
+ ],
+ "angle": 0,
+ "content": "The code and dataset are available| En-Es | Transcription: and always as a child I had this fantasy that somebody would come and rescue me\nMixSpeech: and always as a child I had this fantasy that somebody would come and rescue me |
| Ground Truth: y de niña siempre tenía la fantasía de que alguien vendrá a salvar me de\nMixSpeech: y de (desde) niña niño siempre tenía la esta fantasía de que alguien vendrá a salvar me (rescató) de |
| En-Fr | Transcription: and solutions create new problems which have to be solved in their turn\nMixSpeech: and solutions to create new problems which have to be solved in their turn this year |
| Ground Truth: et les solutions créent de nouveaux problèmes devant être résolus à leur tour\nMixSpeech: et les (des) solutions pour creer de nouveaux problèmes devant qui doivent être résolus à leur tour |
| En-It | Transcription: one of the last 10,000 years, and the other certainly of the last 25 years\nMixSpeech: one of the last 10,000 years and the other certainly assistancee of the last 25 years |
| Ground Truth: uno presente negli ultimi diecimila anni e l'alto certamente negli ultimi 25 anni\nMixSpeech: uno presente negli ultimi diecimila (10000) anni e l'alto certamente negli ultimi 25 anni |
| En-Pt | Transcription: has around 2,000 people descending on MIT's campus\nMixSpeech: has around 2,000 people descending on MIT's campus |
| Ground Truth: congregamerca de 2000 pessoas no campus do mit\nMixSpeech: congregam de circa de 2000 pessoas a estudar no campus do mit |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.076,
+ 0.489,
+ 0.893,
+ 0.518
+ ],
+ "angle": 0,
+ "content": "Table 5. Qualitative performance of Visual Speech Recognition and Translation on AVMuST-TED. Red Strikeout Words: mistranslated words with opposite meaning, (Blue Words in parentheses): mistranslated words with similar meaning, Gray Words: the absent words."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.542,
+ 0.255,
+ 0.558
+ ],
+ "angle": 0,
+ "content": "4.7. Qualitative results"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.569,
+ 0.473,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "We present several examples of lip translation in Table 5 to qualitatively evaluate the translation quality of MixSpeech. The translation results are very close to the ground truth, and the semantics are consistent. We observe two types of words that differ in translation: synonyms and context-sensitive translations. Synonyms that have different spellings but the same meaning, such as salvar and rescató in Spanish, both meaning 'rescue', and diecimila and 10000 in Italian, both meaning 'ten thousand', are commonly found in translation tasks and can affect translation consistency. Additionally, there are translations that require context information, such as when the speaker refers to themselves as a child, and the translation in Spanish needs to take into account the speaker's gender to choose between niña for girl' or Niño for boy' and 'child'. The qualitative translation results of MixSpeech demonstrate its capability to achieve reliable cross-lingua lip translation. In Appendix C.2, we also provide translation results of noisy audio speech translation with visual speech translation and audio speech translation with audiovisual speech translation to highlight the importance of visual speech in speech translation."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.541,
+ 0.619,
+ 0.557
+ ],
+ "angle": 0,
+ "content": "5. Conclusion"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.568,
+ 0.892,
+ 0.794
+ ],
+ "angle": 0,
+ "content": "With the advancement of online technologies, such as online healthcare and sales, language barriers often prevent these tools from reaching and benefiting disadvantaged areas. In light of this, we focus on visual speech, a branch of the speech stream, and aim to translate visual speech from source languages to other target languages for cross-linguistic communication, specifically through lip translation. We meticulously curate the AVMuST-TED dataset, consisting of 706 hours of speech clips with professional translations from TED, to facilitate cross-linguistic research on visual speech. We also introduce MixSpeech, a cross-modality self-learning framework that utilizes mixed speech to regularize visual speech translation and achieves state-of-the-art performance in lip translation on AVMuST-TED and lip reading on LRS2, LRS3, and CMLR datasets."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.795,
+ 0.892,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Moreover, our work on visual speech and AVMuSTED lay a solid foundation for further research on visual speech in cross-lingual fields. There are numerous related tasks with great potential for practical applications, such as Cross-Lingual Talking Head Generation [44]. These tasks hold immense promise for breaking down language barriers and promoting communication across diverse communities."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.519,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "15742"
+ }
+ ],
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.079,
+ 0.091,
+ 0.245,
+ 0.108
+ ],
+ "angle": 0,
+ "content": "Acknowledgements"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.115,
+ 0.47,
+ 0.19
+ ],
+ "angle": 0,
+ "content": "This work was supported in part by National Key R&D Program of China under Grant No.2022ZD0162000, National Natural Science Foundation of China under Grant No.62222211, Grant No.61836002 and Grant No.62072397."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.08,
+ 0.204,
+ 0.173,
+ 0.219
+ ],
+ "angle": 0,
+ "content": "References"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.228,
+ 0.469,
+ 0.284
+ ],
+ "angle": 0,
+ "content": "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 1, 2, 5, 6, 12, 13"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.285,
+ 0.469,
+ 0.34
+ ],
+ "angle": 0,
+ "content": "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 2, 5, 12, 13"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.341,
+ 0.469,
+ 0.41
+ ],
+ "angle": 0,
+ "content": "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.411,
+ 0.469,
+ 0.452
+ ],
+ "angle": 0,
+ "content": "[4] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.453,
+ 0.469,
+ 0.509
+ ],
+ "angle": 0,
+ "content": "[5] Kartik Audhkhasi, Andrew Rosenberg, Abhinav Sethy, Bhuvana Ramabhadran, and Brian Kingsbury. End-to-end asr-free keyword search from speech. IEEE Journal of Selected Topics in Signal Processing, 11(8):1351-1359, 2017. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.51,
+ 0.469,
+ 0.564
+ ],
+ "angle": 0,
+ "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.565,
+ 0.469,
+ 0.619
+ ],
+ "angle": 0,
+ "content": "[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 2, 3, 4"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.621,
+ 0.469,
+ 0.676
+ ],
+ "angle": 0,
+ "content": "[8] Alexandre Bérard, Olivier Pietquin, Christophe Servan, and Laurent Besacier. Listen and translate: A proof of concept for end-to-end speech-to-text translation. arXiv preprint arXiv:1612.01744, 2016. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.087,
+ 0.677,
+ 0.469,
+ 0.745
+ ],
+ "angle": 0,
+ "content": "[9] Peter F Brown, John Cocke, Stephen A Della Pietra, Vincent J Della Pietra, Frederick Jelinek, John Lafferty, Robert L Mercer, and Paul S Roossin. A statistical approach to machine translation. Computational linguistics, 16(2):79-85, 1990. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.747,
+ 0.469,
+ 0.802
+ ],
+ "angle": 0,
+ "content": "[10] Jiaao Chen, Zichao Yang, and Diyi Yang. Mixtext: Linguistically-informed interpolation of hidden space for semi-supervised text classification. arXiv preprint arXiv:2004.12239, 2020. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.804,
+ 0.469,
+ 0.857
+ ],
+ "angle": 0,
+ "content": "[11] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. Opensr: Open-modality speech recognition via maintaining multi-modality alignment. arXiv preprint arXiv:2306.06410, 2023. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.859,
+ 0.469,
+ 0.899
+ ],
+ "angle": 0,
+ "content": "[12] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 13"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.08,
+ 0.228,
+ 0.469,
+ 0.899
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.093,
+ 0.892,
+ 0.162
+ ],
+ "angle": 0,
+ "content": "[13] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 2, 12"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.163,
+ 0.892,
+ 0.219
+ ],
+ "angle": 0,
+ "content": "[14] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. An audio-visual corpus for speech perception and automatic speech recognition. The Journal of the Acoustical Society of America, 120(5):2421-2424, 2006. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.22,
+ 0.892,
+ 0.289
+ ],
+ "angle": 0,
+ "content": "[15] Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672, 2022. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.291,
+ 0.892,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "[16] Mattia A Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. Must-c: a multilingual speech translation corpus. In 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2012-2017. Association for Computational Linguistics, 2019. 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.375,
+ 0.892,
+ 0.432
+ ],
+ "angle": 0,
+ "content": "[17] Qianqian Dong, Fengpeng Yue, Tom Ko, Mingxuan Wang, Qibing Bai, and Yu Zhang. Leveraging pseudo-labeled data to improve direct speech-to-speech translation. arXiv preprint arXiv:2205.08993, 2022. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.433,
+ 0.892,
+ 0.473
+ ],
+ "angle": 0,
+ "content": "[18] Qingkai Fang and Yang Feng. Neural machine translation with phrase-level universal visual representations. arXiv preprint arXiv:2203.10299, 2022. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.475,
+ 0.892,
+ 0.531
+ ],
+ "angle": 0,
+ "content": "[19] Qingkai Fang, Rong Ye, Lei Li, Yang Feng, and Mingxuan Wang. Stemm: Self-learning with speech-text manifold mixup for speech translation. arXiv preprint arXiv:2203.10426, 2022. 1, 2, 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.532,
+ 0.892,
+ 0.588
+ ],
+ "angle": 0,
+ "content": "[20] Hongyu Guo, Yongyi Mao, and Richong Zhang. Mixup as locally linear out-of-manifold regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3714-3722, 2019. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.589,
+ 0.892,
+ 0.63
+ ],
+ "angle": 0,
+ "content": "[21] Naomi Harte and Eoin Gillen. Tcd-limit: An audio-visual corpus of continuous speech. IEEE Transactions on Multimedia, 17(5):603-615, 2015. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.632,
+ 0.892,
+ 0.714
+ ],
+ "angle": 0,
+ "content": "[22] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.715,
+ 0.892,
+ 0.758
+ ],
+ "angle": 0,
+ "content": "[23] Wei-Ning Hsu and Bowen Shi. A single self-supervised model for many speech modalities enables zero-shot modality transfer. arXiv preprint arXiv:2207.07036, 2022.1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.759,
+ 0.892,
+ 0.815
+ ],
+ "angle": 0,
+ "content": "[24] Chenxu Hu, Qiao Tian, Tingle Li, Wang Yuping, Yuxuan Wang, and Hang Zhao. Neural dubber: Dubbing for videos according to scripts. Advances in Neural Information Processing Systems, 34:16582-16595, 2021. 1, 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.816,
+ 0.892,
+ 0.871
+ ],
+ "angle": 0,
+ "content": "[25] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.872,
+ 0.892,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "[26] Rongjie Huang, Huadai Liu, Xize Cheng, Yi Ren, Linjun Li, Zhenhui Ye, Jinzheng He, Lichao Zhang, Jinglin Liu, Xiang"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.503,
+ 0.093,
+ 0.892,
+ 0.902
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.519,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "15743"
+ }
+ ],
+ [
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.108,
+ 0.092,
+ 0.47,
+ 0.133
+ ],
+ "angle": 0,
+ "content": "Yin, et al. Av-transpeech: Audio-visual robust speech-to-speech translation. arXiv preprint arXiv:2305.15403, 2023. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.134,
+ 0.47,
+ 0.189
+ ],
+ "angle": 0,
+ "content": "[27] Rongjie Huang, Yi Ren, Jinglin Liu, Chenye Cui, and Zhou Zhao. Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech. In Advances in Neural Information Processing Systems. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.19,
+ 0.47,
+ 0.258
+ ],
+ "angle": 0,
+ "content": "[28] Rongjie Huang, Zhou Zhao, Huadai Liu, Jinglin Liu, Chenye Cui, and Yi Ren. Prodiff: Progressive fast diffusion model for high-quality text-to-speech. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2595-2605, 2022. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.26,
+ 0.47,
+ 0.314
+ ],
+ "angle": 0,
+ "content": "[29] Rongjie Huang, Zhou Zhao, Jinglin Liu, Huadai Liu, Yi Ren, Lichao Zhang, and Jinzheng He. Transpeech: Speech-to-speech translation with bilateral perturbation. arXiv preprint arXiv:2205.12523, 2022. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.316,
+ 0.47,
+ 0.357
+ ],
+ "angle": 0,
+ "content": "[30] Taejun Kim and Juhan Nam. Temporal feedback convolutional recurrent neural networks for keyword spotting. arXiv preprint arXiv:1911.01803, 2019. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.359,
+ 0.47,
+ 0.398
+ ],
+ "angle": 0,
+ "content": "[31] Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 12"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.4,
+ 0.47,
+ 0.481
+ ],
+ "angle": 0,
+ "content": "[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.484,
+ 0.47,
+ 0.538
+ ],
+ "angle": 0,
+ "content": "[33] Ann Lee, Peng-Jen Chen, Changhan Wang, Jiatao Gu, Xutai Ma, Adam Polyak, Yossi Adi, Qing He, Yun Tang, Juan Pino, et al. Direct speech-to-speech translation with discrete units. arXiv preprint arXiv:2107.05604, 2021. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.539,
+ 0.47,
+ 0.607
+ ],
+ "angle": 0,
+ "content": "[34] Linjun Li, Tao Jin, Xize Cheng, Ye Wang, Wang Lin, Rongjie Huang, and Zhou Zhao. Contrastive token-wise meta-learning for unseen performer visual temporal-aligned translation. In Findings of the Association for Computational Linguistics: ACL 2023, pages 10993-11007, 2023. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.609,
+ 0.47,
+ 0.65
+ ],
+ "angle": 0,
+ "content": "[35] Rainer Lienhart. Reliable transition detection in videos: A survey and practitioner's guide. International journal of image and graphics, 1(03):469-486, 2001. 12"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.651,
+ 0.47,
+ 0.719
+ ],
+ "angle": 0,
+ "content": "[36] Zhijie Lin, Zhou Zhao, Haoyuan Li, Jinglin Liu, Meng Zhang, Xingshan Zeng, and Xiaofei He. Simullr: Simultaneous lip reading transducer with attention-guided adaptive memory. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1359-1367, 2021. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.721,
+ 0.47,
+ 0.789
+ ],
+ "angle": 0,
+ "content": "[37] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.79,
+ 0.47,
+ 0.845
+ ],
+ "angle": 0,
+ "content": "[38] Yuchen Liu, Hao Xiong, Zhongjun He, Jiajun Zhang, Hua Wu, Haifeng Wang, and Chengqing Zong. End-to-end speech translation with knowledge distillation. arXiv preprint arXiv:1904.08075, 2019. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.846,
+ 0.47,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "[39] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. Lira: Learning visual speech representations from audio through self-supervision. arXiv preprint arXiv:2106.09171, 2021. 2, 12"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.08,
+ 0.092,
+ 0.47,
+ 0.901
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.093,
+ 0.892,
+ 0.147
+ ],
+ "angle": 0,
+ "content": "[40] Shihui Ma, Shilin Wang, and Xiang Lin. A transformer-based model for sentence-level chinese mandarin lipreading. In 2020 IEEE Fifth International Conference on Data Science in Cyberspace (DSC), pages 78-81. IEEE, 2020. 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.149,
+ 0.892,
+ 0.189
+ ],
+ "angle": 0,
+ "content": "[41] ML Menéndez, JA Pardo, L Pardo, and MC Pardo. The jensen-shannon divergence. Journal of the Franklin Institute, 334(2):307-318, 1997. 4"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.19,
+ 0.892,
+ 0.245
+ ],
+ "angle": 0,
+ "content": "[42] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. arXiv preprint arXiv:2009.01225, 2020. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.247,
+ 0.892,
+ 0.274
+ ],
+ "angle": 0,
+ "content": "[43] David S Moore. Uncertainty. On the shoulders of giants: New approaches to numeracy, pages 95-137, 1990. 3, 4"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.275,
+ 0.892,
+ 0.302
+ ],
+ "angle": 0,
+ "content": "[44] NVIDIA. Nvidia maxine: Reinventing real-time video communications with ai. 2022. 1, 2, 8"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.303,
+ 0.892,
+ 0.37
+ ],
+ "angle": 0,
+ "content": "[45] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 2, 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.372,
+ 0.892,
+ 0.427
+ ],
+ "angle": 0,
+ "content": "[46] Matt Post. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Belgium, Brussels, Oct. 2018. Association for Computational Linguistics. 5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.428,
+ 0.892,
+ 0.483
+ ],
+ "angle": 0,
+ "content": "[47] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 2, 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.484,
+ 0.892,
+ 0.552
+ ],
+ "angle": 0,
+ "content": "[48] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.554,
+ 0.892,
+ 0.635
+ ],
+ "angle": 0,
+ "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.637,
+ 0.892,
+ 0.706
+ ],
+ "angle": 0,
+ "content": "[50] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.707,
+ 0.892,
+ 0.762
+ ],
+ "angle": 0,
+ "content": "[51] Max Ritter, Uwe Meier, Jie Yang, and Alex Waibel. Face translation: A multimodal translation agent. In AVSP'99-International Conference on Auditory-Visual Speech Processing, 1999. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.763,
+ 0.892,
+ 0.843
+ ],
+ "angle": 0,
+ "content": "[52] Andrew Rosenberg, Kartik Audhkhasi, Abhinav Sethy, Bhuvana Ramabhadran, and Michael Picheny. End-to-end speech recognition and keyword search on low-resource languages. In 2017 IEEE international conference on acoustics, speech and signal processing (icassp), pages 5280-5284. IEEE, 2017. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.845,
+ 0.892,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "[53] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. arXiv preprint arXiv:2201.02184, 2022. 1, 2, 3, 4, 5, 6, 12, 13"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.503,
+ 0.093,
+ 0.892,
+ 0.901
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.52,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "15744"
+ }
+ ],
+ [
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.08,
+ 0.092,
+ 0.468,
+ 0.134
+ ],
+ "angle": 0,
+ "content": "[54] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 1, 13"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.137,
+ 0.468,
+ 0.178
+ ],
+ "angle": 0,
+ "content": "[55] David Snyder, Guoguo Chen, and Daniel Povey. Musan: A music, speech, and noise corpus. ArXiv, abs/1510.08484, 2015. 6, 13"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.182,
+ 0.468,
+ 0.236
+ ],
+ "angle": 0,
+ "content": "[56] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.24,
+ 0.468,
+ 0.322
+ ],
+ "angle": 0,
+ "content": "[57] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21425–21430, 2022. 1, 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.326,
+ 0.468,
+ 0.368
+ ],
+ "angle": 0,
+ "content": "[58] Matthias Sperber, Graham Neubig, Jan Niehues, and Alex Waibel. Neural lattice-to-sequence models for uncertain inputs. arXiv preprint arXiv:1704.00559, 2017. 1, 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.371,
+ 0.468,
+ 0.452
+ ],
+ "angle": 0,
+ "content": "[59] Yun Tang, Juan Pino, Changhan Wang, Xutai Ma, and Dmitriy Genzel. A general multi-task learning framework to leverage text data for speech to text tasks. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6209-6213. IEEE, 2021. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.457,
+ 0.468,
+ 0.539
+ ],
+ "angle": 0,
+ "content": "[60] Alexander Waibel, Moritz Behr, Fevziye Irem Eyiokur, Dogucan Yaman, Tuan-Nam Nguyen, Carlos Mullov, Mehmet Arif Demirtas, Alperen Kantarcı, Stefan Constantin, and Hazim Kemal Ekenel. Face-dubbing++: Lipsynchronous, voice preserving translation of videos. arXiv preprint arXiv:2206.04523, 2022. 1, 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.543,
+ 0.468,
+ 0.598
+ ],
+ "angle": 0,
+ "content": "[61] Chen Wang, Yuchen Liu, Boxing Chen, Jiajun Zhang, Wei Luo, Zhongqiang Huang, and Chengqing Zong. Discrete cross-modal alignment enables zero-shot speech translation. arXiv preprint arXiv:2210.09556, 2022. 1"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.601,
+ 0.468,
+ 0.683
+ ],
+ "angle": 0,
+ "content": "[62] Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. arXiv preprint arXiv:2101.00390, 2021.5"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.687,
+ 0.468,
+ 0.755
+ ],
+ "angle": 0,
+ "content": "[63] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.759,
+ 0.468,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "[64] Zehan Wang, Yang Zhao, Xize Cheng, Haifeng Huang, Jiageng Liu, Li Tang, Linjun Li, Yongqi Wang, Aoxiong Yin, Zhang Zhang, et al. Connecting multi-modal contrastive representations. arXiv preprint arXiv:2305.14381, 2023. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.081,
+ 0.818,
+ 0.468,
+ 0.9
+ ],
+ "angle": 0,
+ "content": "[65] Shuang Yang, Yuanhang Zhang, Dalu Feng, Mingmin Yang, Chenhao Wang, Jingyun Xiao, Keyu Long, Shiguang Shan, and Xilin Chen. Lrw-1000: A naturally-distributed large-scale benchmark for lip reading in the wild. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1–8. IEEE, 2019. 12"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.08,
+ 0.092,
+ 0.468,
+ 0.9
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.503,
+ 0.093,
+ 0.892,
+ 0.134
+ ],
+ "angle": 0,
+ "content": "[66] Rong Ye, Mingxuan Wang, and Lei Li. End-to-end speech translation via cross-modal progressive training. arXiv preprint arXiv:2104.10380, 2021. 2"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.137,
+ 0.892,
+ 0.176
+ ],
+ "angle": 0,
+ "content": "[67] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.179,
+ 0.892,
+ 0.232
+ ],
+ "angle": 0,
+ "content": "[68] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, Amirata Ghorbani, and James Zou. How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819, 2020. 3"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.235,
+ 0.892,
+ 0.29
+ ],
+ "angle": 0,
+ "content": "[69] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.292,
+ 0.892,
+ 0.345
+ ],
+ "angle": 0,
+ "content": "[70] Ya Zhao, Rui Xu, and Mingli Song. A cascade sequence-to-sequence model for chinese mandarin lip reading. In Proceedings of the ACM Multimedia Asia, pages 1-6. 2019. 5, 6"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.504,
+ 0.348,
+ 0.892,
+ 0.416
+ ],
+ "angle": 0,
+ "content": "[71] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6917-6924, 2020. 2, 5, 6"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.503,
+ 0.093,
+ 0.892,
+ 0.416
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.946,
+ 0.519,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "15745"
+ }
+ ]
+]
\ No newline at end of file
diff --git a/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/8115edb8-8391-4f40-bf06-0cd9421711d5_origin.pdf b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/8115edb8-8391-4f40-bf06-0cd9421711d5_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..ab4ce4f47ae32446641dd203d2d01671bd8b3da0
--- /dev/null
+++ b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/8115edb8-8391-4f40-bf06-0cd9421711d5_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d2fa4f5932f550b8e359a9a8ac761e7b6734b2a1393cfa388ddd328afe425aa
+size 1408727
diff --git a/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/full.md b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..86aa7df5b3ab03688422803e26585ec2679abfae
--- /dev/null
+++ b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/full.md
@@ -0,0 +1,294 @@
+# MixSpeech: Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition
+
+Xize Cheng*, Tao Jin*, Rongjie Huang*, Linjun Li
+Wang Lin, Zehan Wang, Ye Wang, Huadai Liu, Aoxiong Yin, Zhou Zhao†
+Zhejiang University
+{chengxize, jint.zju, rongjiehuang, zhaozhou}@zju.edu.cn
+
+# Abstract
+
+Multi-media communications facilitate global interaction among people. However, despite researchers exploring cross-lingual translation techniques such as machine translation and audio speech translation to overcome language barriers, there is still a shortage of cross-lingual studies on visual speech. This lack of research is mainly due to the absence of datasets containing visual speech and translated text pairs. In this paper, we present AVMuST-TED, the first dataset for Audio-Visual Multilingual Speech Translation, derived from TED talks. Nonetheless, visual speech is not as distinguishable as audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this issue, we propose MixSpeech, a cross-modality self-learning framework that utilizes audio speech to regularize the training of visual speech tasks. To further minimize the cross-modality gap and its impact on knowledge transfer, we suggest adopting mixed speech, which is created by interpolating audio and visual streams, along with a curriculum learning strategy to adjust the mixing ratio as needed. MixSpeech enhances speech translation in noisy environments, improving BLEU scores for four languages on AVMuST-TED by $+1.4$ to $+4.2$ . Moreover, it achieves state-of-the-art performance in lip reading on CMLR $(11.1\%)$ , LRS2 $(25.5\%)$ , and LRS3 $(28.0\%)$ .
+
+# 1. Introduction
+
+Multi-media techniques, including Audio-Visual Speech Recognition (AVSR) [4, 1, 2, 53], Audio-Visual Speech Translation (AVST) [8, 38, 61], and Audio-Visual Speech Generation (AVSG) [48, 32, 24], are commonly employed in various online communication scenarios, such as conferences, education, and healthcare. As a tool for ultra-remote communication, many online interactions involve multiple
+
+
+Figure 1. Diagram of speech tasks. Audio speech and visual speech are paired parallel speech streams which can be employed for speech recognition and speech translation. However, only Lip-Translation remains unexplored.
+
+En: Hello, beautiful world.
+
+Es: Hola, hermoso mundo.
+
+Fr: Bonjour, beau monde.
+
+It : Ciao, bel mondo.
+
+Pt: Olá, belo mundo.
+
+···
+
+languages, prompting the need for addressing cross-lingual challenges. Several works have attempted to tackle these challenges, including Machine Translation (MT) [9, 37, 15] for text utterance, Speech Translation (ST) [58, 19] for audio utterance, and Speech-to-Speech Translation (S2ST) [58, 19, 17, 33, 29] for simultaneous interpretation. However, research on cross-lingual visual speech is still limited, as illustrated in Figure 1. As an essential component of multi-media speech, visual speech can be combined with audio to enhance the recognition and understanding of speech content as audio-visual speech [1, 2, 54], and is the unique resource for speech content understanding in audio-disabled scenarios [36].
+
+Visual speech translation has never been studied, mainly for the lack of visual speech datasets with translated texts in different languages. The few remaining works [57, 60, 44] also cannot be quantitatively verified for this reason, making them unconvincing. The available visual speech corpus is often very scarce compared to audio speech owing to the high demands of visual speech for model training, which requires mostly-frontal and high-resolution videos with a sufficiently high frame rate, such that motions around the lip area are clearly captured [23]. In this paper, we propose the
+
+first Audio-Visual Multilingual Speech Translation dataset, AVMuST-TED. During the process of acquisition, we first screen out videos with professional translations in four different languages from TED talk which performs strict translation and review processes, and then determine the real speaker's talking head by checking whether each pair of visual speech (i.e., talking head) and audio speech matches in the manner of [1, 2]. Incidentally, this dataset can also be used for quantitative evaluation of other multi-modality translation tasks, such as cross-lingua audio-visual speech generation [51, 60].
+
+The cascaded model comprising of a speech recognition model and a machine translation model can handle speech translation tasks but suffers from error accumulation due to model cascades and cannot process languages without text (e.g., Minnan). Our proposed end-to-end model, which can translate directly from source speech to target text, addresses the above issues. However, visual speech is less distinguishable than audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this, we introduce MixSpeech, a method that first pretrains the decoder using high-discrimination audio speech to obtain a mapping from speech phonemes to text and then generalizes this mapping to the visual speech task through cross-modality self-learning. Furthermore, since audio speech and visual speech are two distinct modalities of speech, there is a significant modality gap between them that hinders knowledge transfer. To narrow this gap and improve knowledge transfer, we propose mixed speech, which is created by interpolating audio and visual streams, rather than relying solely on audio speech. We also propose a curriculum-learning [7] based strategy to adjust the mixing ratio as the training progresses and cross-modality integration deepens.
+
+The code and dataset are available| En-Es | Transcription: and always as a child I had this fantasy that somebody would come and rescue me
+MixSpeech: and always as a child I had this fantasy that somebody would come and rescue me |
| Ground Truth: y de niña siempre tenía la fantasía de que alguien vendrá a salvar me de
+MixSpeech: y de (desde) niña niño siempre tenía la esta fantasía de que alguien vendrá a salvar me (rescató) de |
| En-Fr | Transcription: and solutions create new problems which have to be solved in their turn
+MixSpeech: and solutions to create new problems which have to be solved in their turn this year |
| Ground Truth: et les solutions créent de nouveaux problèmes devant être résolus à leur tour
+MixSpeech: et les (des) solutions pour creer de nouveaux problèmes devant qui doivent être résolus à leur tour |
| En-It | Transcription: one of the last 10,000 years, and the other certainly of the last 25 years
+MixSpeech: one of the last 10,000 years and the other certainly assistancee of the last 25 years |
| Ground Truth: uno presente negli ultimi diecimila anni e l'alto certamente negli ultimi 25 anni
+MixSpeech: uno presente negli ultimi diecimila (10000) anni e l'alto certamente negli ultimi 25 anni |
| En-Pt | Transcription: has around 2,000 people descending on MIT's campus
+MixSpeech: has around 2,000 people descending on MIT's campus |
| Ground Truth: congregamerca de 2000 pessoas no campus do mit
+MixSpeech: congregam de circa de 2000 pessoas a estudar no campus do mit |
+
+Table 5. Qualitative performance of Visual Speech Recognition and Translation on AVMuST-TED. Red Strikeout Words: mistranslated words with opposite meaning, (Blue Words in parentheses): mistranslated words with similar meaning, Gray Words: the absent words.
+
+# 4.7. Qualitative results
+
+We present several examples of lip translation in Table 5 to qualitatively evaluate the translation quality of MixSpeech. The translation results are very close to the ground truth, and the semantics are consistent. We observe two types of words that differ in translation: synonyms and context-sensitive translations. Synonyms that have different spellings but the same meaning, such as salvar and rescató in Spanish, both meaning 'rescue', and diecimila and 10000 in Italian, both meaning 'ten thousand', are commonly found in translation tasks and can affect translation consistency. Additionally, there are translations that require context information, such as when the speaker refers to themselves as a child, and the translation in Spanish needs to take into account the speaker's gender to choose between niña for girl' or Niño for boy' and 'child'. The qualitative translation results of MixSpeech demonstrate its capability to achieve reliable cross-lingua lip translation. In Appendix C.2, we also provide translation results of noisy audio speech translation with visual speech translation and audio speech translation with audiovisual speech translation to highlight the importance of visual speech in speech translation.
+
+# 5. Conclusion
+
+With the advancement of online technologies, such as online healthcare and sales, language barriers often prevent these tools from reaching and benefiting disadvantaged areas. In light of this, we focus on visual speech, a branch of the speech stream, and aim to translate visual speech from source languages to other target languages for cross-linguistic communication, specifically through lip translation. We meticulously curate the AVMuST-TED dataset, consisting of 706 hours of speech clips with professional translations from TED, to facilitate cross-linguistic research on visual speech. We also introduce MixSpeech, a cross-modality self-learning framework that utilizes mixed speech to regularize visual speech translation and achieves state-of-the-art performance in lip translation on AVMuST-TED and lip reading on LRS2, LRS3, and CMLR datasets.
+
+Moreover, our work on visual speech and AVMuSTED lay a solid foundation for further research on visual speech in cross-lingual fields. There are numerous related tasks with great potential for practical applications, such as Cross-Lingual Talking Head Generation [44]. These tasks hold immense promise for breaking down language barriers and promoting communication across diverse communities.
+
+# Acknowledgements
+
+This work was supported in part by National Key R&D Program of China under Grant No.2022ZD0162000, National Natural Science Foundation of China under Grant No.62222211, Grant No.61836002 and Grant No.62072397.
+
+# References
+
+[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 1, 2, 5, 6, 12, 13
+[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 2, 5, 12, 13
+[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 6
+[4] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 1
+[5] Kartik Audhkhasi, Andrew Rosenberg, Abhinav Sethy, Bhuvana Ramabhadran, and Brian Kingsbury. End-to-end asr-free keyword search from speech. IEEE Journal of Selected Topics in Signal Processing, 11(8):1351-1359, 2017. 2
+[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 2
+[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 2, 3, 4
+[8] Alexandre Bérard, Olivier Pietquin, Christophe Servan, and Laurent Besacier. Listen and translate: A proof of concept for end-to-end speech-to-text translation. arXiv preprint arXiv:1612.01744, 2016. 1
+[9] Peter F Brown, John Cocke, Stephen A Della Pietra, Vincent J Della Pietra, Frederick Jelinek, John Lafferty, Robert L Mercer, and Paul S Roossin. A statistical approach to machine translation. Computational linguistics, 16(2):79-85, 1990. 1
+[10] Jiaao Chen, Zichao Yang, and Diyi Yang. Mixtext: Linguistically-informed interpolation of hidden space for semi-supervised text classification. arXiv preprint arXiv:2004.12239, 2020. 3
+[11] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. Opensr: Open-modality speech recognition via maintaining multi-modality alignment. arXiv preprint arXiv:2306.06410, 2023. 2
+[12] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 13
+
+[13] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 2, 12
+[14] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. An audio-visual corpus for speech perception and automatic speech recognition. The Journal of the Acoustical Society of America, 120(5):2421-2424, 2006. 2
+[15] Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672, 2022. 1
+[16] Mattia A Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. Must-c: a multilingual speech translation corpus. In 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2012-2017. Association for Computational Linguistics, 2019. 5
+[17] Qianqian Dong, Fengpeng Yue, Tom Ko, Mingxuan Wang, Qibing Bai, and Yu Zhang. Leveraging pseudo-labeled data to improve direct speech-to-speech translation. arXiv preprint arXiv:2205.08993, 2022. 1
+[18] Qingkai Fang and Yang Feng. Neural machine translation with phrase-level universal visual representations. arXiv preprint arXiv:2203.10299, 2022. 3
+[19] Qingkai Fang, Rong Ye, Lei Li, Yang Feng, and Mingxuan Wang. Stemm: Self-learning with speech-text manifold mixup for speech translation. arXiv preprint arXiv:2203.10426, 2022. 1, 2, 3
+[20] Hongyu Guo, Yongyi Mao, and Richong Zhang. Mixup as locally linear out-of-manifold regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3714-3722, 2019. 3
+[21] Naomi Harte and Eoin Gillen. Tcd-limit: An audio-visual corpus of continuous speech. IEEE Transactions on Multimedia, 17(5):603-615, 2015. 2
+[22] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 2
+[23] Wei-Ning Hsu and Bowen Shi. A single self-supervised model for many speech modalities enables zero-shot modality transfer. arXiv preprint arXiv:2207.07036, 2022.1
+[24] Chenxu Hu, Qiao Tian, Tingle Li, Wang Yuping, Yuxuan Wang, and Hang Zhao. Neural dubber: Dubbing for videos according to scripts. Advances in Neural Information Processing Systems, 34:16582-16595, 2021. 1, 3
+[25] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022. 2
+[26] Rongjie Huang, Huadai Liu, Xize Cheng, Yi Ren, Linjun Li, Zhenhui Ye, Jinzheng He, Lichao Zhang, Jinglin Liu, Xiang
+
+Yin, et al. Av-transpeech: Audio-visual robust speech-to-speech translation. arXiv preprint arXiv:2305.15403, 2023. 2
+[27] Rongjie Huang, Yi Ren, Jinglin Liu, Chenye Cui, and Zhou Zhao. Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech. In Advances in Neural Information Processing Systems. 3
+[28] Rongjie Huang, Zhou Zhao, Huadai Liu, Jinglin Liu, Chenye Cui, and Yi Ren. Prodiff: Progressive fast diffusion model for high-quality text-to-speech. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2595-2605, 2022. 2
+[29] Rongjie Huang, Zhou Zhao, Jinglin Liu, Huadai Liu, Yi Ren, Lichao Zhang, and Jinzheng He. Transpeech: Speech-to-speech translation with bilateral perturbation. arXiv preprint arXiv:2205.12523, 2022. 1
+[30] Taejun Kim and Juhan Nam. Temporal feedback convolutional recurrent neural networks for keyword spotting. arXiv preprint arXiv:1911.01803, 2019. 2
+[31] Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 12
+[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 1
+[33] Ann Lee, Peng-Jen Chen, Changhan Wang, Jiatao Gu, Xutai Ma, Adam Polyak, Yossi Adi, Qing He, Yun Tang, Juan Pino, et al. Direct speech-to-speech translation with discrete units. arXiv preprint arXiv:2107.05604, 2021. 1
+[34] Linjun Li, Tao Jin, Xize Cheng, Ye Wang, Wang Lin, Rongjie Huang, and Zhou Zhao. Contrastive token-wise meta-learning for unseen performer visual temporal-aligned translation. In Findings of the Association for Computational Linguistics: ACL 2023, pages 10993-11007, 2023. 2
+[35] Rainer Lienhart. Reliable transition detection in videos: A survey and practitioner's guide. International journal of image and graphics, 1(03):469-486, 2001. 12
+[36] Zhijie Lin, Zhou Zhao, Haoyuan Li, Jinglin Liu, Meng Zhang, Xingshan Zeng, and Xiaofei He. Simullr: Simultaneous lip reading transducer with attention-guided adaptive memory. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1359-1367, 2021. 1
+[37] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 1
+[38] Yuchen Liu, Hao Xiong, Zhongjun He, Jiajun Zhang, Hua Wu, Haifeng Wang, and Chengqing Zong. End-to-end speech translation with knowledge distillation. arXiv preprint arXiv:1904.08075, 2019. 1
+[39] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. Lira: Learning visual speech representations from audio through self-supervision. arXiv preprint arXiv:2106.09171, 2021. 2, 12
+
+[40] Shihui Ma, Shilin Wang, and Xiang Lin. A transformer-based model for sentence-level chinese mandarin lipreading. In 2020 IEEE Fifth International Conference on Data Science in Cyberspace (DSC), pages 78-81. IEEE, 2020. 6
+[41] ML Menéndez, JA Pardo, L Pardo, and MC Pardo. The jensen-shannon divergence. Journal of the Franklin Institute, 334(2):307-318, 1997. 4
+[42] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. arXiv preprint arXiv:2009.01225, 2020. 2
+[43] David S Moore. Uncertainty. On the shoulders of giants: New approaches to numeracy, pages 95-137, 1990. 3, 4
+[44] NVIDIA. Nvidia maxine: Reinventing real-time video communications with ai. 2022. 1, 2, 8
+[45] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 2, 5
+[46] Matt Post. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Belgium, Brussels, Oct. 2018. Association for Computational Linguistics. 5
+[47] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 2, 6
+[48] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1
+[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3
+[50] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6
+[51] Max Ritter, Uwe Meier, Jie Yang, and Alex Waibel. Face translation: A multimodal translation agent. In AVSP'99-International Conference on Auditory-Visual Speech Processing, 1999. 2
+[52] Andrew Rosenberg, Kartik Audhkhasi, Abhinav Sethy, Bhuvana Ramabhadran, and Michael Picheny. End-to-end speech recognition and keyword search on low-resource languages. In 2017 IEEE international conference on acoustics, speech and signal processing (icassp), pages 5280-5284. IEEE, 2017. 2
+[53] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. arXiv preprint arXiv:2201.02184, 2022. 1, 2, 3, 4, 5, 6, 12, 13
+
+[54] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 1, 13
+[55] David Snyder, Guoguo Chen, and Daniel Povey. Musan: A music, speech, and noise corpus. ArXiv, abs/1510.08484, 2015. 6, 13
+[56] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 6
+[57] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21425–21430, 2022. 1, 2
+[58] Matthias Sperber, Graham Neubig, Jan Niehues, and Alex Waibel. Neural lattice-to-sequence models for uncertain inputs. arXiv preprint arXiv:1704.00559, 2017. 1, 2
+[59] Yun Tang, Juan Pino, Changhan Wang, Xutai Ma, and Dmitriy Genzel. A general multi-task learning framework to leverage text data for speech to text tasks. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6209-6213. IEEE, 2021. 3
+[60] Alexander Waibel, Moritz Behr, Fevziye Irem Eyiokur, Dogucan Yaman, Tuan-Nam Nguyen, Carlos Mullov, Mehmet Arif Demirtas, Alperen Kantarcı, Stefan Constantin, and Hazim Kemal Ekenel. Face-dubbing++: Lipsynchronous, voice preserving translation of videos. arXiv preprint arXiv:2206.04523, 2022. 1, 2
+[61] Chen Wang, Yuchen Liu, Boxing Chen, Jiajun Zhang, Wei Luo, Zhongqiang Huang, and Chengqing Zong. Discrete cross-modal alignment enables zero-shot speech translation. arXiv preprint arXiv:2210.09556, 2022. 1
+[62] Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. arXiv preprint arXiv:2101.00390, 2021.5
+[63] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 3
+[64] Zehan Wang, Yang Zhao, Xize Cheng, Haifeng Huang, Jiageng Liu, Li Tang, Linjun Li, Yongqi Wang, Aoxiong Yin, Zhang Zhang, et al. Connecting multi-modal contrastive representations. arXiv preprint arXiv:2305.14381, 2023. 2
+[65] Shuang Yang, Yuanhang Zhang, Dalu Feng, Mingmin Yang, Chenhao Wang, Jingyun Xiao, Keyu Long, Shiguang Shan, and Xilin Chen. Lrw-1000: A naturally-distributed large-scale benchmark for lip reading in the wild. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1–8. IEEE, 2019. 12
+
+[66] Rong Ye, Mingxuan Wang, and Lei Li. End-to-end speech translation via cross-modal progressive training. arXiv preprint arXiv:2104.10380, 2021. 2
+[67] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3
+[68] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, Amirata Ghorbani, and James Zou. How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819, 2020. 3
+[69] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 6
+[70] Ya Zhao, Rui Xu, and Mingli Song. A cascade sequence-to-sequence model for chinese mandarin lip reading. In Proceedings of the ACM Multimedia Asia, pages 1-6. 2019. 5, 6
+[71] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6917-6924, 2020. 2, 5, 6
\ No newline at end of file
diff --git a/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/images.zip b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/images.zip
new file mode 100644
index 0000000000000000000000000000000000000000..686b50d6c9ae6d914faabebef8a4c39fefc5e3a4
--- /dev/null
+++ b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/images.zip
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b54d728bf01fab52a3bd02680780b03ad5850350f9d5407087ac6df21898e8ed
+size 620058
diff --git a/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/layout.json b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..9c29705f40c3b16ab8c38616fe0de4ec675d1fc0
--- /dev/null
+++ b/2023/MixSpeech_ Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition/layout.json
@@ -0,0 +1,8288 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 62,
+ 102,
+ 532,
+ 140
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 62,
+ 102,
+ 532,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 62,
+ 102,
+ 532,
+ 140
+ ],
+ "type": "text",
+ "content": "MixSpeech: Cross-Modality Self-Learning with Audio-Visual Stream Mixup for Visual Speech Translation and Recognition"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 112,
+ 161,
+ 485,
+ 217
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 161,
+ 485,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 161,
+ 485,
+ 217
+ ],
+ "type": "text",
+ "content": "Xize Cheng*, Tao Jin*, Rongjie Huang*, Linjun Li \nWang Lin, Zehan Wang, Ye Wang, Huadai Liu, Aoxiong Yin, Zhou Zhao† \nZhejiang University \n{chengxize, jint.zju, rongjiehuang, zhaozhou}@zju.edu.cn"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 143,
+ 244,
+ 192,
+ 258
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 244,
+ 192,
+ 258
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 244,
+ 192,
+ 258
+ ],
+ "type": "text",
+ "content": "Abstract"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "text",
+ "content": "Multi-media communications facilitate global interaction among people. However, despite researchers exploring cross-lingual translation techniques such as machine translation and audio speech translation to overcome language barriers, there is still a shortage of cross-lingual studies on visual speech. This lack of research is mainly due to the absence of datasets containing visual speech and translated text pairs. In this paper, we present AVMuST-TED, the first dataset for Audio-Visual Multilingual Speech Translation, derived from TED talks. Nonetheless, visual speech is not as distinguishable as audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this issue, we propose MixSpeech, a cross-modality self-learning framework that utilizes audio speech to regularize the training of visual speech tasks. To further minimize the cross-modality gap and its impact on knowledge transfer, we suggest adopting mixed speech, which is created by interpolating audio and visual streams, along with a curriculum learning strategy to adjust the mixing ratio as needed. MixSpeech enhances speech translation in noisy environments, improving BLEU scores for four languages on AVMuST-TED by "
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "inline_equation",
+ "content": "+1.4"
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "text",
+ "content": " to "
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "inline_equation",
+ "content": "+4.2"
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "text",
+ "content": ". Moreover, it achieves state-of-the-art performance in lip reading on CMLR "
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "inline_equation",
+ "content": "(11.1\\%)"
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "text",
+ "content": ", LRS2 "
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "inline_equation",
+ "content": "(25.5\\%)"
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "text",
+ "content": ", and LRS3 "
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "inline_equation",
+ "content": "(28.0\\%)"
+ },
+ {
+ "bbox": [
+ 47,
+ 270,
+ 290,
+ 558
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 47,
+ 581,
+ 128,
+ 594
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 581,
+ 128,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 581,
+ 128,
+ 594
+ ],
+ "type": "text",
+ "content": "1. Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 46,
+ 601,
+ 287,
+ 685
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 601,
+ 287,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 601,
+ 287,
+ 685
+ ],
+ "type": "text",
+ "content": "Multi-media techniques, including Audio-Visual Speech Recognition (AVSR) [4, 1, 2, 53], Audio-Visual Speech Translation (AVST) [8, 38, 61], and Audio-Visual Speech Generation (AVSG) [48, 32, 24], are commonly employed in various online communication scenarios, such as conferences, education, and healthcare. As a tool for ultra-remote communication, many online interactions involve multiple"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 312,
+ 258,
+ 451,
+ 355
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 312,
+ 258,
+ 451,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 312,
+ 258,
+ 451,
+ 355
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 312,
+ 258,
+ 451,
+ 355
+ ],
+ "type": "image",
+ "image_path": "9e335c4f0355b0c07d0ce08b0bcae93eb7596c8a0e96b9b4080f4064649ec260.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 304,
+ 365,
+ 547,
+ 410
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 365,
+ 547,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 365,
+ 547,
+ 410
+ ],
+ "type": "text",
+ "content": "Figure 1. Diagram of speech tasks. Audio speech and visual speech are paired parallel speech streams which can be employed for speech recognition and speech translation. However, only Lip-Translation remains unexplored."
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 457,
+ 264,
+ 537,
+ 274
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 457,
+ 264,
+ 537,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 457,
+ 264,
+ 537,
+ 274
+ ],
+ "type": "text",
+ "content": "En: Hello, beautiful world."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 457,
+ 299,
+ 539,
+ 308
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 457,
+ 299,
+ 539,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 457,
+ 299,
+ 539,
+ 308
+ ],
+ "type": "text",
+ "content": "Es: Hola, hermoso mundo."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 457,
+ 312,
+ 539,
+ 320
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 457,
+ 312,
+ 539,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 457,
+ 312,
+ 539,
+ 320
+ ],
+ "type": "text",
+ "content": "Fr: Bonjour, beau monde."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 457,
+ 323,
+ 524,
+ 333
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 457,
+ 323,
+ 524,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 457,
+ 323,
+ 524,
+ 333
+ ],
+ "type": "text",
+ "content": "It : Ciao, bel mondo."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 457,
+ 336,
+ 524,
+ 346
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 457,
+ 336,
+ 524,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 457,
+ 336,
+ 524,
+ 346
+ ],
+ "type": "text",
+ "content": "Pt: Olá, belo mundo."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 457,
+ 348,
+ 479,
+ 354
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 457,
+ 348,
+ 479,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 457,
+ 348,
+ 479,
+ 354
+ ],
+ "type": "text",
+ "content": "···"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 304,
+ 435,
+ 547,
+ 590
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 435,
+ 547,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 435,
+ 547,
+ 590
+ ],
+ "type": "text",
+ "content": "languages, prompting the need for addressing cross-lingual challenges. Several works have attempted to tackle these challenges, including Machine Translation (MT) [9, 37, 15] for text utterance, Speech Translation (ST) [58, 19] for audio utterance, and Speech-to-Speech Translation (S2ST) [58, 19, 17, 33, 29] for simultaneous interpretation. However, research on cross-lingual visual speech is still limited, as illustrated in Figure 1. As an essential component of multi-media speech, visual speech can be combined with audio to enhance the recognition and understanding of speech content as audio-visual speech [1, 2, 54], and is the unique resource for speech content understanding in audio-disabled scenarios [36]."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 304,
+ 594,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 594,
+ 548,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 594,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": "Visual speech translation has never been studied, mainly for the lack of visual speech datasets with translated texts in different languages. The few remaining works [57, 60, 44] also cannot be quantitatively verified for this reason, making them unconvincing. The available visual speech corpus is often very scarce compared to audio speech owing to the high demands of visual speech for model training, which requires mostly-frontal and high-resolution videos with a sufficiently high frame rate, such that motions around the lip area are clearly captured [23]. In this paper, we propose the"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 65,
+ 2,
+ 111,
+ 34
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 65,
+ 2,
+ 111,
+ 34
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 65,
+ 2,
+ 111,
+ 34
+ ],
+ "type": "text",
+ "content": "CVF"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 146,
+ 0,
+ 494,
+ 37
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 0,
+ 494,
+ 37
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 0,
+ 494,
+ 37
+ ],
+ "type": "text",
+ "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 58,
+ 693,
+ 127,
+ 703
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 58,
+ 693,
+ 127,
+ 703
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 58,
+ 693,
+ 127,
+ 703
+ ],
+ "type": "text",
+ "content": "*Equal contribution."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 59,
+ 703,
+ 135,
+ 712
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 59,
+ 703,
+ 135,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 59,
+ 703,
+ 135,
+ 712
+ ],
+ "type": "text",
+ "content": "† Corresponding author."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "type": "text",
+ "content": "15735"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 46,
+ 72,
+ 289,
+ 203
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 72,
+ 289,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 72,
+ 289,
+ 203
+ ],
+ "type": "text",
+ "content": "first Audio-Visual Multilingual Speech Translation dataset, AVMuST-TED. During the process of acquisition, we first screen out videos with professional translations in four different languages from TED talk which performs strict translation and review processes, and then determine the real speaker's talking head by checking whether each pair of visual speech (i.e., talking head) and audio speech matches in the manner of [1, 2]. Incidentally, this dataset can also be used for quantitative evaluation of other multi-modality translation tasks, such as cross-lingua audio-visual speech generation [51, 60]."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 46,
+ 205,
+ 289,
+ 479
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 205,
+ 289,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 205,
+ 289,
+ 479
+ ],
+ "type": "text",
+ "content": "The cascaded model comprising of a speech recognition model and a machine translation model can handle speech translation tasks but suffers from error accumulation due to model cascades and cannot process languages without text (e.g., Minnan). Our proposed end-to-end model, which can translate directly from source speech to target text, addresses the above issues. However, visual speech is less distinguishable than audio speech, making it difficult to develop a mapping from source speech phonemes to the target language text. To address this, we introduce MixSpeech, a method that first pretrains the decoder using high-discrimination audio speech to obtain a mapping from speech phonemes to text and then generalizes this mapping to the visual speech task through cross-modality self-learning. Furthermore, since audio speech and visual speech are two distinct modalities of speech, there is a significant modality gap between them that hinders knowledge transfer. To narrow this gap and improve knowledge transfer, we propose mixed speech, which is created by interpolating audio and visual streams, rather than relying solely on audio speech. We also propose a curriculum-learning [7] based strategy to adjust the mixing ratio as the training progresses and cross-modality integration deepens."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 47,
+ 480,
+ 287,
+ 506
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 480,
+ 287,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 480,
+ 287,
+ 506
+ ],
+ "type": "text",
+ "content": "The code and dataset are available| En-Es | Transcription: and always as a child I had this fantasy that somebody would come and rescue me\nMixSpeech: and always as a child I had this fantasy that somebody would come and rescue me |
| Ground Truth: y de niña siempre tenía la fantasía de que alguien vendrá a salvar me de\nMixSpeech: y de (desde) niña niño siempre tenía la esta fantasía de que alguien vendrá a salvar me (rescató) de |
| En-Fr | Transcription: and solutions create new problems which have to be solved in their turn\nMixSpeech: and solutions to create new problems which have to be solved in their turn this year |
| Ground Truth: et les solutions créent de nouveaux problèmes devant être résolus à leur tour\nMixSpeech: et les (des) solutions pour creer de nouveaux problèmes devant qui doivent être résolus à leur tour |
| En-It | Transcription: one of the last 10,000 years, and the other certainly of the last 25 years\nMixSpeech: one of the last 10,000 years and the other certainly assistancee of the last 25 years |
| Ground Truth: uno presente negli ultimi diecimila anni e l'alto certamente negli ultimi 25 anni\nMixSpeech: uno presente negli ultimi diecimila (10000) anni e l'alto certamente negli ultimi 25 anni |
| En-Pt | Transcription: has around 2,000 people descending on MIT's campus\nMixSpeech: has around 2,000 people descending on MIT's campus |
| Ground Truth: congregamerca de 2000 pessoas no campus do mit\nMixSpeech: congregam de circa de 2000 pessoas a estudar no campus do mit |
",
+ "image_path": "29845b784ea69738221c57fbdadb394d9accb123fcee2958b33841140671eac9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 46,
+ 387,
+ 546,
+ 410
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 387,
+ 546,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 387,
+ 546,
+ 410
+ ],
+ "type": "text",
+ "content": "Table 5. Qualitative performance of Visual Speech Recognition and Translation on AVMuST-TED. Red Strikeout Words: mistranslated words with opposite meaning, (Blue Words in parentheses): mistranslated words with similar meaning, Gray Words: the absent words."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 47,
+ 429,
+ 156,
+ 441
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 429,
+ 156,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 429,
+ 156,
+ 441
+ ],
+ "type": "text",
+ "content": "4.7. Qualitative results"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 46,
+ 450,
+ 289,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 450,
+ 289,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 450,
+ 289,
+ 715
+ ],
+ "type": "text",
+ "content": "We present several examples of lip translation in Table 5 to qualitatively evaluate the translation quality of MixSpeech. The translation results are very close to the ground truth, and the semantics are consistent. We observe two types of words that differ in translation: synonyms and context-sensitive translations. Synonyms that have different spellings but the same meaning, such as salvar and rescató in Spanish, both meaning 'rescue', and diecimila and 10000 in Italian, both meaning 'ten thousand', are commonly found in translation tasks and can affect translation consistency. Additionally, there are translations that require context information, such as when the speaker refers to themselves as a child, and the translation in Spanish needs to take into account the speaker's gender to choose between niña for girl' or Niño for boy' and 'child'. The qualitative translation results of MixSpeech demonstrate its capability to achieve reliable cross-lingua lip translation. In Appendix C.2, we also provide translation results of noisy audio speech translation with visual speech translation and audio speech translation with audiovisual speech translation to highlight the importance of visual speech in speech translation."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 306,
+ 428,
+ 378,
+ 441
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 428,
+ 378,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 428,
+ 378,
+ 441
+ ],
+ "type": "text",
+ "content": "5. Conclusion"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 304,
+ 449,
+ 545,
+ 628
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 449,
+ 545,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 449,
+ 545,
+ 628
+ ],
+ "type": "text",
+ "content": "With the advancement of online technologies, such as online healthcare and sales, language barriers often prevent these tools from reaching and benefiting disadvantaged areas. In light of this, we focus on visual speech, a branch of the speech stream, and aim to translate visual speech from source languages to other target languages for cross-linguistic communication, specifically through lip translation. We meticulously curate the AVMuST-TED dataset, consisting of 706 hours of speech clips with professional translations from TED, to facilitate cross-linguistic research on visual speech. We also introduce MixSpeech, a cross-modality self-learning framework that utilizes mixed speech to regularize visual speech translation and achieves state-of-the-art performance in lip translation on AVMuST-TED and lip reading on LRS2, LRS3, and CMLR datasets."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 629,
+ 545,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 629,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": "Moreover, our work on visual speech and AVMuSTED lay a solid foundation for further research on visual speech in cross-lingual fields. There are numerous related tasks with great potential for practical applications, such as Cross-Lingual Talking Head Generation [44]. These tasks hold immense promise for breaking down language barriers and promoting communication across diverse communities."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "type": "text",
+ "content": "15742"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 7
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 149,
+ 85
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 149,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 149,
+ 85
+ ],
+ "type": "text",
+ "content": "Acknowledgements"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 47,
+ 91,
+ 287,
+ 150
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 91,
+ 287,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 91,
+ 287,
+ 150
+ ],
+ "type": "text",
+ "content": "This work was supported in part by National Key R&D Program of China under Grant No.2022ZD0162000, National Natural Science Foundation of China under Grant No.62222211, Grant No.61836002 and Grant No.62072397."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 48,
+ 161,
+ 105,
+ 173
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 161,
+ 105,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 161,
+ 105,
+ 173
+ ],
+ "type": "text",
+ "content": "References"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 48,
+ 180,
+ 287,
+ 712
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 15,
+ "blocks": [
+ {
+ "bbox": [
+ 53,
+ 180,
+ 287,
+ 224
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 180,
+ 287,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 180,
+ 287,
+ 224
+ ],
+ "type": "text",
+ "content": "[1] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 1, 2, 5, 6, 12, 13"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 53,
+ 225,
+ 287,
+ 269
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 225,
+ 287,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 225,
+ 287,
+ 269
+ ],
+ "type": "text",
+ "content": "[2] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 2, 5, 12, 13"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 53,
+ 270,
+ 287,
+ 324
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 270,
+ 287,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 270,
+ 287,
+ 324
+ ],
+ "type": "text",
+ "content": "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 6"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 53,
+ 325,
+ 287,
+ 357
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 325,
+ 287,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 325,
+ 287,
+ 357
+ ],
+ "type": "text",
+ "content": "[4] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 1"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 53,
+ 358,
+ 287,
+ 403
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 358,
+ 287,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 358,
+ 287,
+ 403
+ ],
+ "type": "text",
+ "content": "[5] Kartik Audhkhasi, Andrew Rosenberg, Abhinav Sethy, Bhuvana Ramabhadran, and Brian Kingsbury. End-to-end asr-free keyword search from speech. IEEE Journal of Selected Topics in Signal Processing, 11(8):1351-1359, 2017. 2"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 53,
+ 403,
+ 287,
+ 446
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 403,
+ 287,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 403,
+ 287,
+ 446
+ ],
+ "type": "text",
+ "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 2"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 53,
+ 447,
+ 287,
+ 490
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 447,
+ 287,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 447,
+ 287,
+ 490
+ ],
+ "type": "text",
+ "content": "[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. 2, 3, 4"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 53,
+ 491,
+ 287,
+ 535
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 491,
+ 287,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 491,
+ 287,
+ 535
+ ],
+ "type": "text",
+ "content": "[8] Alexandre Bérard, Olivier Pietquin, Christophe Servan, and Laurent Besacier. Listen and translate: A proof of concept for end-to-end speech-to-text translation. arXiv preprint arXiv:1612.01744, 2016. 1"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 53,
+ 536,
+ 287,
+ 590
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 53,
+ 536,
+ 287,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 53,
+ 536,
+ 287,
+ 590
+ ],
+ "type": "text",
+ "content": "[9] Peter F Brown, John Cocke, Stephen A Della Pietra, Vincent J Della Pietra, Frederick Jelinek, John Lafferty, Robert L Mercer, and Paul S Roossin. A statistical approach to machine translation. Computational linguistics, 16(2):79-85, 1990. 1"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 48,
+ 591,
+ 287,
+ 635
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 591,
+ 287,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 591,
+ 287,
+ 635
+ ],
+ "type": "text",
+ "content": "[10] Jiaao Chen, Zichao Yang, and Diyi Yang. Mixtext: Linguistically-informed interpolation of hidden space for semi-supervised text classification. arXiv preprint arXiv:2004.12239, 2020. 3"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 48,
+ 636,
+ 287,
+ 678
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 636,
+ 287,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 636,
+ 287,
+ 678
+ ],
+ "type": "text",
+ "content": "[11] Xize Cheng, Tao Jin, Linjun Li, Wang Lin, Xinyu Duan, and Zhou Zhao. Opensr: Open-modality speech recognition via maintaining multi-modality alignment. arXiv preprint arXiv:2306.06410, 2023. 2"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 48,
+ 680,
+ 287,
+ 712
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 680,
+ 287,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 680,
+ 287,
+ 712
+ ],
+ "type": "text",
+ "content": "[12] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. In INTERSPEECH, 2018. 13"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 714
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 30,
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 128
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 128
+ ],
+ "type": "text",
+ "content": "[13] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 2, 12"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 307,
+ 129,
+ 545,
+ 173
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 129,
+ 545,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 129,
+ 545,
+ 173
+ ],
+ "type": "text",
+ "content": "[14] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. An audio-visual corpus for speech perception and automatic speech recognition. The Journal of the Acoustical Society of America, 120(5):2421-2424, 2006. 2"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 307,
+ 174,
+ 545,
+ 228
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 174,
+ 545,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 174,
+ 545,
+ 228
+ ],
+ "type": "text",
+ "content": "[15] Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672, 2022. 1"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 308,
+ 230,
+ 545,
+ 296
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 230,
+ 545,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 230,
+ 545,
+ 296
+ ],
+ "type": "text",
+ "content": "[16] Mattia A Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. Must-c: a multilingual speech translation corpus. In 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2012-2017. Association for Computational Linguistics, 2019. 5"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 308,
+ 297,
+ 545,
+ 342
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 297,
+ 545,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 297,
+ 545,
+ 342
+ ],
+ "type": "text",
+ "content": "[17] Qianqian Dong, Fengpeng Yue, Tom Ko, Mingxuan Wang, Qibing Bai, and Yu Zhang. Leveraging pseudo-labeled data to improve direct speech-to-speech translation. arXiv preprint arXiv:2205.08993, 2022. 1"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 308,
+ 342,
+ 545,
+ 374
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 342,
+ 545,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 342,
+ 545,
+ 374
+ ],
+ "type": "text",
+ "content": "[18] Qingkai Fang and Yang Feng. Neural machine translation with phrase-level universal visual representations. arXiv preprint arXiv:2203.10299, 2022. 3"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 308,
+ 376,
+ 545,
+ 420
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 376,
+ 545,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 376,
+ 545,
+ 420
+ ],
+ "type": "text",
+ "content": "[19] Qingkai Fang, Rong Ye, Lei Li, Yang Feng, and Mingxuan Wang. Stemm: Self-learning with speech-text manifold mixup for speech translation. arXiv preprint arXiv:2203.10426, 2022. 1, 2, 3"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 308,
+ 421,
+ 545,
+ 465
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 421,
+ 545,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 421,
+ 545,
+ 465
+ ],
+ "type": "text",
+ "content": "[20] Hongyu Guo, Yongyi Mao, and Richong Zhang. Mixup as locally linear out-of-manifold regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3714-3722, 2019. 3"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 308,
+ 466,
+ 545,
+ 498
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 466,
+ 545,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 466,
+ 545,
+ 498
+ ],
+ "type": "text",
+ "content": "[21] Naomi Harte and Eoin Gillen. Tcd-limit: An audio-visual corpus of continuous speech. IEEE Transactions on Multimedia, 17(5):603-615, 2015. 2"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 308,
+ 500,
+ 545,
+ 565
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 500,
+ 545,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 500,
+ 545,
+ 565
+ ],
+ "type": "text",
+ "content": "[22] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451-3460, 2021. 2"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 308,
+ 566,
+ 545,
+ 600
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 566,
+ 545,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 566,
+ 545,
+ 600
+ ],
+ "type": "text",
+ "content": "[23] Wei-Ning Hsu and Bowen Shi. A single self-supervised model for many speech modalities enables zero-shot modality transfer. arXiv preprint arXiv:2207.07036, 2022.1"
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 308,
+ 601,
+ 545,
+ 645
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 601,
+ 545,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 601,
+ 545,
+ 645
+ ],
+ "type": "text",
+ "content": "[24] Chenxu Hu, Qiao Tian, Tingle Li, Wang Yuping, Yuxuan Wang, and Hang Zhao. Neural dubber: Dubbing for videos according to scripts. Advances in Neural Information Processing Systems, 34:16582-16595, 2021. 1, 3"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 308,
+ 646,
+ 545,
+ 689
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 646,
+ 545,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 646,
+ 545,
+ 689
+ ],
+ "type": "text",
+ "content": "[25] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022. 2"
+ }
+ ]
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 308,
+ 690,
+ 545,
+ 714
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 690,
+ 545,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 690,
+ 545,
+ 714
+ ],
+ "type": "text",
+ "content": "[26] Rongjie Huang, Huadai Liu, Xize Cheng, Yi Ren, Linjun Li, Zhenhui Ye, Jinzheng He, Lichao Zhang, Jinglin Liu, Xiang"
+ }
+ ]
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 317,
+ 757
+ ],
+ "type": "text",
+ "content": "15743"
+ }
+ ]
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 8
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 287,
+ 713
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 14,
+ "blocks": [
+ {
+ "bbox": [
+ 66,
+ 72,
+ 287,
+ 105
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 66,
+ 72,
+ 287,
+ 105
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 66,
+ 72,
+ 287,
+ 105
+ ],
+ "type": "text",
+ "content": "Yin, et al. Av-transpeech: Audio-visual robust speech-to-speech translation. arXiv preprint arXiv:2305.15403, 2023. 2"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 48,
+ 106,
+ 287,
+ 149
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 106,
+ 287,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 106,
+ 287,
+ 149
+ ],
+ "type": "text",
+ "content": "[27] Rongjie Huang, Yi Ren, Jinglin Liu, Chenye Cui, and Zhou Zhao. Generspeech: Towards style transfer for generalizable out-of-domain text-to-speech. In Advances in Neural Information Processing Systems. 3"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 48,
+ 150,
+ 287,
+ 204
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 150,
+ 287,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 150,
+ 287,
+ 204
+ ],
+ "type": "text",
+ "content": "[28] Rongjie Huang, Zhou Zhao, Huadai Liu, Jinglin Liu, Chenye Cui, and Yi Ren. Prodiff: Progressive fast diffusion model for high-quality text-to-speech. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2595-2605, 2022. 2"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 48,
+ 205,
+ 287,
+ 248
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 205,
+ 287,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 205,
+ 287,
+ 248
+ ],
+ "type": "text",
+ "content": "[29] Rongjie Huang, Zhou Zhao, Jinglin Liu, Huadai Liu, Yi Ren, Lichao Zhang, and Jinzheng He. Transpeech: Speech-to-speech translation with bilateral perturbation. arXiv preprint arXiv:2205.12523, 2022. 1"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 48,
+ 250,
+ 287,
+ 282
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 250,
+ 287,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 250,
+ 287,
+ 282
+ ],
+ "type": "text",
+ "content": "[30] Taejun Kim and Juhan Nam. Temporal feedback convolutional recurrent neural networks for keyword spotting. arXiv preprint arXiv:1911.01803, 2019. 2"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 48,
+ 284,
+ 287,
+ 315
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 284,
+ 287,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 284,
+ 287,
+ 315
+ ],
+ "type": "text",
+ "content": "[31] Davis E King. Dlib-ml: A machine learning toolkit. The Journal of Machine Learning Research, 10:1755-1758, 2009. 12"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 48,
+ 316,
+ 287,
+ 380
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 316,
+ 287,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 316,
+ 287,
+ 380
+ ],
+ "type": "text",
+ "content": "[32] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 1"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 48,
+ 383,
+ 287,
+ 426
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 383,
+ 287,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 383,
+ 287,
+ 426
+ ],
+ "type": "text",
+ "content": "[33] Ann Lee, Peng-Jen Chen, Changhan Wang, Jiatao Gu, Xutai Ma, Adam Polyak, Yossi Adi, Qing He, Yun Tang, Juan Pino, et al. Direct speech-to-speech translation with discrete units. arXiv preprint arXiv:2107.05604, 2021. 1"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 48,
+ 426,
+ 287,
+ 480
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 426,
+ 287,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 426,
+ 287,
+ 480
+ ],
+ "type": "text",
+ "content": "[34] Linjun Li, Tao Jin, Xize Cheng, Ye Wang, Wang Lin, Rongjie Huang, and Zhou Zhao. Contrastive token-wise meta-learning for unseen performer visual temporal-aligned translation. In Findings of the Association for Computational Linguistics: ACL 2023, pages 10993-11007, 2023. 2"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 48,
+ 482,
+ 287,
+ 514
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 482,
+ 287,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 482,
+ 287,
+ 514
+ ],
+ "type": "text",
+ "content": "[35] Rainer Lienhart. Reliable transition detection in videos: A survey and practitioner's guide. International journal of image and graphics, 1(03):469-486, 2001. 12"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 48,
+ 515,
+ 287,
+ 569
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 515,
+ 287,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 515,
+ 287,
+ 569
+ ],
+ "type": "text",
+ "content": "[36] Zhijie Lin, Zhou Zhao, Haoyuan Li, Jinglin Liu, Meng Zhang, Xingshan Zeng, and Xiaofei He. Simullr: Simultaneous lip reading transducer with attention-guided adaptive memory. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1359-1367, 2021. 1"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 48,
+ 571,
+ 287,
+ 624
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 571,
+ 287,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 571,
+ 287,
+ 624
+ ],
+ "type": "text",
+ "content": "[37] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 1"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 48,
+ 625,
+ 287,
+ 669
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 625,
+ 287,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 625,
+ 287,
+ 669
+ ],
+ "type": "text",
+ "content": "[38] Yuchen Liu, Hao Xiong, Zhongjun He, Jiajun Zhang, Hua Wu, Haifeng Wang, and Chengqing Zong. End-to-end speech translation with knowledge distillation. arXiv preprint arXiv:1904.08075, 2019. 1"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 48,
+ 670,
+ 287,
+ 713
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 670,
+ 287,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 670,
+ 287,
+ 713
+ ],
+ "type": "text",
+ "content": "[39] Pingchuan Ma, Rodrigo Mira, Stavros Petridis, Bjorn W Schuller, and Maja Pantic. Lira: Learning visual speech representations from audio through self-supervision. arXiv preprint arXiv:2106.09171, 2021. 2, 12"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 713
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 29,
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 116
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 116
+ ],
+ "type": "text",
+ "content": "[40] Shihui Ma, Shilin Wang, and Xiang Lin. A transformer-based model for sentence-level chinese mandarin lipreading. In 2020 IEEE Fifth International Conference on Data Science in Cyberspace (DSC), pages 78-81. IEEE, 2020. 6"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 307,
+ 118,
+ 545,
+ 149
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 118,
+ 545,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 118,
+ 545,
+ 149
+ ],
+ "type": "text",
+ "content": "[41] ML Menéndez, JA Pardo, L Pardo, and MC Pardo. The jensen-shannon divergence. Journal of the Franklin Institute, 334(2):307-318, 1997. 4"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 307,
+ 150,
+ 545,
+ 194
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 150,
+ 545,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 150,
+ 545,
+ 194
+ ],
+ "type": "text",
+ "content": "[42] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. arXiv preprint arXiv:2009.01225, 2020. 2"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 307,
+ 195,
+ 545,
+ 217
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 195,
+ 545,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 195,
+ 545,
+ 217
+ ],
+ "type": "text",
+ "content": "[43] David S Moore. Uncertainty. On the shoulders of giants: New approaches to numeracy, pages 95-137, 1990. 3, 4"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 307,
+ 217,
+ 545,
+ 239
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 217,
+ 545,
+ 239
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 217,
+ 545,
+ 239
+ ],
+ "type": "text",
+ "content": "[44] NVIDIA. Nvidia maxine: Reinventing real-time video communications with ai. 2022. 1, 2, 8"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 307,
+ 239,
+ 545,
+ 293
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 239,
+ 545,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 239,
+ 545,
+ 293
+ ],
+ "type": "text",
+ "content": "[45] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 2, 5"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 307,
+ 294,
+ 545,
+ 338
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 294,
+ 545,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 294,
+ 545,
+ 338
+ ],
+ "type": "text",
+ "content": "[46] Matt Post. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Belgium, Brussels, Oct. 2018. Association for Computational Linguistics. 5"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 307,
+ 338,
+ 545,
+ 382
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 338,
+ 545,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 338,
+ 545,
+ 382
+ ],
+ "type": "text",
+ "content": "[47] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5162-5172, 2022. 2, 6"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 307,
+ 383,
+ 545,
+ 437
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 383,
+ 545,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 383,
+ 545,
+ 437
+ ],
+ "type": "text",
+ "content": "[48] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 1"
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 307,
+ 438,
+ 545,
+ 502
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 438,
+ 545,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 438,
+ 545,
+ 502
+ ],
+ "type": "text",
+ "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 307,
+ 504,
+ 545,
+ 559
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 504,
+ 545,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 504,
+ 545,
+ 559
+ ],
+ "type": "text",
+ "content": "[50] Sucheng Ren, Yong Du, Jianming Lv, Guoqiang Han, and Shengfeng He. Learning from the master: Distilling cross-modal advanced knowledge for lip reading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13325-13333, 2021. 2, 6"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 307,
+ 559,
+ 545,
+ 603
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 559,
+ 545,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 559,
+ 545,
+ 603
+ ],
+ "type": "text",
+ "content": "[51] Max Ritter, Uwe Meier, Jie Yang, and Alex Waibel. Face translation: A multimodal translation agent. In AVSP'99-International Conference on Auditory-Visual Speech Processing, 1999. 2"
+ }
+ ]
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 307,
+ 604,
+ 545,
+ 667
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 604,
+ 545,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 604,
+ 545,
+ 667
+ ],
+ "type": "text",
+ "content": "[52] Andrew Rosenberg, Kartik Audhkhasi, Abhinav Sethy, Bhuvana Ramabhadran, and Michael Picheny. End-to-end speech recognition and keyword search on low-resource languages. In 2017 IEEE international conference on acoustics, speech and signal processing (icassp), pages 5280-5284. IEEE, 2017. 2"
+ }
+ ]
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 307,
+ 669,
+ 545,
+ 713
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 669,
+ 545,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 669,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": "[53] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. arXiv preprint arXiv:2201.02184, 2022. 1, 2, 3, 4, 5, 6, 12, 13"
+ }
+ ]
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 318,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 318,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 318,
+ 757
+ ],
+ "type": "text",
+ "content": "15744"
+ }
+ ]
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 9
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 286,
+ 712
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 12,
+ "blocks": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 286,
+ 106
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 286,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 72,
+ 286,
+ 106
+ ],
+ "type": "text",
+ "content": "[54] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 1, 13"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 49,
+ 108,
+ 286,
+ 140
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 108,
+ 286,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 108,
+ 286,
+ 140
+ ],
+ "type": "text",
+ "content": "[55] David Snyder, Guoguo Chen, and Daniel Povey. Musan: A music, speech, and noise corpus. ArXiv, abs/1510.08484, 2015. 6, 13"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 49,
+ 144,
+ 286,
+ 186
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 144,
+ 286,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 144,
+ 286,
+ 186
+ ],
+ "type": "text",
+ "content": "[56] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 6"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 49,
+ 190,
+ 286,
+ 255
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 190,
+ 286,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 190,
+ 286,
+ 255
+ ],
+ "type": "text",
+ "content": "[57] Hyoung-Kyu Song, Sang Hoon Woo, Junhyeok Lee, Seung-min Yang, Hyunjae Cho, Youseong Lee, Dongho Choi, and Kang-wook Kim. Talking face generation with multilingual tts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21425–21430, 2022. 1, 2"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 49,
+ 258,
+ 286,
+ 291
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 258,
+ 286,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 258,
+ 286,
+ 291
+ ],
+ "type": "text",
+ "content": "[58] Matthias Sperber, Graham Neubig, Jan Niehues, and Alex Waibel. Neural lattice-to-sequence models for uncertain inputs. arXiv preprint arXiv:1704.00559, 2017. 1, 2"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 49,
+ 293,
+ 286,
+ 357
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 293,
+ 286,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 293,
+ 286,
+ 357
+ ],
+ "type": "text",
+ "content": "[59] Yun Tang, Juan Pino, Changhan Wang, Xutai Ma, and Dmitriy Genzel. A general multi-task learning framework to leverage text data for speech to text tasks. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6209-6213. IEEE, 2021. 3"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 49,
+ 361,
+ 286,
+ 426
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 361,
+ 286,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 361,
+ 286,
+ 426
+ ],
+ "type": "text",
+ "content": "[60] Alexander Waibel, Moritz Behr, Fevziye Irem Eyiokur, Dogucan Yaman, Tuan-Nam Nguyen, Carlos Mullov, Mehmet Arif Demirtas, Alperen Kantarcı, Stefan Constantin, and Hazim Kemal Ekenel. Face-dubbing++: Lipsynchronous, voice preserving translation of videos. arXiv preprint arXiv:2206.04523, 2022. 1, 2"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 49,
+ 430,
+ 286,
+ 473
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 430,
+ 286,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 430,
+ 286,
+ 473
+ ],
+ "type": "text",
+ "content": "[61] Chen Wang, Yuchen Liu, Boxing Chen, Jiajun Zhang, Wei Luo, Zhongqiang Huang, and Chengqing Zong. Discrete cross-modal alignment enables zero-shot speech translation. arXiv preprint arXiv:2210.09556, 2022. 1"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 49,
+ 475,
+ 286,
+ 540
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 475,
+ 286,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 475,
+ 286,
+ 540
+ ],
+ "type": "text",
+ "content": "[62] Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. Voxpopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. arXiv preprint arXiv:2101.00390, 2021.5"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 49,
+ 544,
+ 286,
+ 597
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 544,
+ 286,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 544,
+ 286,
+ 597
+ ],
+ "type": "text",
+ "content": "[63] Teng Wang, Wenhao Jiang, Zhichao Lu, Feng Zheng, Ran Cheng, Chengguo Yin, and Ping Luo. Vlmixer: Unpaired vision-language pre-training via cross-modal cutmix. In International Conference on Machine Learning, pages 22680-22690. PMLR, 2022. 3"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 49,
+ 601,
+ 286,
+ 644
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 601,
+ 286,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 601,
+ 286,
+ 644
+ ],
+ "type": "text",
+ "content": "[64] Zehan Wang, Yang Zhao, Xize Cheng, Haifeng Huang, Jiageng Liu, Li Tang, Linjun Li, Yongqi Wang, Aoxiong Yin, Zhang Zhang, et al. Connecting multi-modal contrastive representations. arXiv preprint arXiv:2305.14381, 2023. 2"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 49,
+ 647,
+ 286,
+ 712
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 647,
+ 286,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 647,
+ 286,
+ 712
+ ],
+ "type": "text",
+ "content": "[65] Shuang Yang, Yuanhang Zhang, Dalu Feng, Mingmin Yang, Chenhao Wang, Jingyun Xiao, Keyu Long, Shiguang Shan, and Xilin Chen. Lrw-1000: A naturally-distributed large-scale benchmark for lip reading in the wild. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1–8. IEEE, 2019. 12"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "sub_type": "ref_text"
+ },
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 329
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 19,
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 106
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 73,
+ 545,
+ 106
+ ],
+ "type": "text",
+ "content": "[66] Rong Ye, Mingxuan Wang, and Lei Li. End-to-end speech translation via cross-modal progressive training. arXiv preprint arXiv:2104.10380, 2021. 2"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 308,
+ 108,
+ 545,
+ 139
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 108,
+ 545,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 108,
+ 545,
+ 139
+ ],
+ "type": "text",
+ "content": "[67] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 308,
+ 141,
+ 545,
+ 183
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 141,
+ 545,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 141,
+ 545,
+ 183
+ ],
+ "type": "text",
+ "content": "[68] Linjun Zhang, Zhun Deng, Kenji Kawaguchi, Amirata Ghorbani, and James Zou. How does mixup help with robustness and generalization? arXiv preprint arXiv:2010.04819, 2020. 3"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 308,
+ 186,
+ 545,
+ 229
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 186,
+ 545,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 186,
+ 545,
+ 229
+ ],
+ "type": "text",
+ "content": "[69] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 713-722, 2019. 6"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 308,
+ 231,
+ 545,
+ 273
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 231,
+ 545,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 231,
+ 545,
+ 273
+ ],
+ "type": "text",
+ "content": "[70] Ya Zhao, Rui Xu, and Mingli Song. A cascade sequence-to-sequence model for chinese mandarin lip reading. In Proceedings of the ACM Multimedia Asia, pages 1-6. 2019. 5, 6"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 308,
+ 275,
+ 545,
+ 329
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 545,
+ 329
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 275,
+ 545,
+ 329
+ ],
+ "type": "text",
+ "content": "[71] Ya Zhao, Rui Xu, Xinchao Wang, Peng Hou, Haihong Tang, and Mingli Song. Hearing lips: Improving lip reading by distilling speech recognizers. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6917-6924, 2020. 2, 5, 6"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 294,
+ 749,
+ 317,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 749,
+ 317,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 749,
+ 317,
+ 757
+ ],
+ "type": "text",
+ "content": "15745"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 10
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2023/MixSynthFormer_ A Transformer Encoder-like Structure with Mixed Synthetic Self-attention for Efficient Human Pose Estimation/846dbe02-0b5b-4bf4-a391-96d12e988697_content_list.json b/2023/MixSynthFormer_ A Transformer Encoder-like Structure with Mixed Synthetic Self-attention for Efficient Human Pose Estimation/846dbe02-0b5b-4bf4-a391-96d12e988697_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..411d412fc1808e2b4200e5cd48ab6d37b9006a46
--- /dev/null
+++ b/2023/MixSynthFormer_ A Transformer Encoder-like Structure with Mixed Synthetic Self-attention for Efficient Human Pose Estimation/846dbe02-0b5b-4bf4-a391-96d12e988697_content_list.json
@@ -0,0 +1,1746 @@
+[
+ {
+ "type": "text",
+ "text": "MixSynthFormer: A Transformer Encoder-like Structure with Mixed Synthetic Self-attention for Efficient Human Pose Estimation",
+ "text_level": 1,
+ "bbox": [
+ 84,
+ 130,
+ 883,
+ 174
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Yuran Sun, Alan William Dougherty, Zhuoying Zhang, Yi King Choi, Chuan Wu The University of Hong Kong",
+ "bbox": [
+ 165,
+ 203,
+ 805,
+ 239
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "yuransun@connect.hku.hk,{alanwd,tinawork}@hku.hk,{ykchoi,cwu}@cs.hku.hk",
+ "bbox": [
+ 171,
+ 241,
+ 795,
+ 257
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract",
+ "text_level": 1,
+ "bbox": [
+ 233,
+ 291,
+ 313,
+ 306
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Human pose estimation in videos has wide-ranging practical applications across various fields, many of which require fast inference on resource-scarce devices, necessitating the development of efficient and accurate algorithms. Previous works have demonstrated the feasibility of exploiting motion continuity to conduct pose estimation using sparsely sampled frames with transformer-based models. However, these methods only consider the temporal relation while neglecting spatial attention, and the complexity of dot product self-attention calculations in transformers are quadratically proportional to the embedding size. To address these limitations, we propose MixSynthFormer, a transformer encoder-like model with MLP-based mixed synthetic attention. By mixing synthesized spatial and temporal attentions, our model incorporates inter-joint and inter-frame importance and can accurately estimate human poses in an entire video sequence from sparsely sampled frames. Additionally, the flexible design of our model makes it versatile for other motion synthesis tasks. Our extensive experiments on 2D/3D pose estimation, body mesh recovery, and motion prediction validate the effectiveness and efficiency of MixSynthFormer. The code is available at https://github.com/irenesesun/MixSynthFormer.git",
+ "bbox": [
+ 75,
+ 325,
+ 473,
+ 688
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1. Introduction",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 722,
+ 209,
+ 738
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Human pose estimation is a crucial task in computer vision which aims to estimate the keypoint locations of a human body from visual inputs. It has a wide range of applications in virtual/augmented reality, healthcare, and security surveillance [23, 5]. In recent years, there has been a growing demand for real-time pose estimation in many of these applications. However, the computational burden of deep learning models makes their real-time estimation on resource-constrained devices (e.g., CPU, mobile devices) challenging. How to estimate human poses in videos ef",
+ "bbox": [
+ 75,
+ 750,
+ 472,
+ 902
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/e8bd756ea35c49dc38310397c312f3996facd74f6e24c1b8465e944151b606af.jpg",
+ "image_caption": [
+ "(a) Feature-based Keyframe Selector"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 517,
+ 290,
+ 877,
+ 340
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/63913df22c7377e0faec804750e093645a5dc651c436b520cf7faf4b3659db39.jpg",
+ "image_caption": [
+ "(b) Sampling-based Keyframe Selector"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 537,
+ 368,
+ 856,
+ 412
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "image",
+ "img_path": "images/7521788d817b6f1540170364e75a89536f202e321c2d7eb40ddba898157e6dd7.jpg",
+ "image_caption": [
+ "(c) Ours",
+ "Figure 1: Pipelines of keyframe-based pose estimation frameworks. The recover process is similar to interpolation. (a) is the feature-based pipeline [40, 7], which selects $30\\%$ to $40\\%$ of frames based on input features and recovers the whole sequence. (b) is the existing sampling-based pipeline [38], which uses a sampler to select frames and conducts keyframe refinement, whole sequence recovery and whole sequence refinement. (c) is our proposed simplified sampling-based pipeline excluding keyframe refinement."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 557,
+ 435,
+ 836,
+ 479
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "frequently remains an open research problem.",
+ "bbox": [
+ 500,
+ 628,
+ 790,
+ 643
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Conventional pose estimators [3, 22, 41, 29, 4, 37, 27] estimate poses frame-by-frame, and the efficiency depends on the model design. Recent methods [40, 7, 38] have proven the viability of estimating poses in sequences based solely on keyframes. By leveraging the temporal redundancy in videos, poses in the remaining frames can be reconstructed from keyframe poses, similar to motion completion, which can significantly reduce the cost of resource-intensive pose estimators.",
+ "bbox": [
+ 496,
+ 643,
+ 893,
+ 777
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "In current keyframe-based pose estimation frameworks, keyframes are selected using a feature-based selector or a sampler, as illustrated in Figures 1a and 1b. Feature-based keyframe selectors [40, 7] select \"good\" frames based on features extracted from inputs. The final poses in the whole sequence are recovered from the estimated poses in those frames without further refinements. A more effective approach is to sample a portion of frames directly from inputs",
+ "bbox": [
+ 496,
+ 779,
+ 895,
+ 902
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "header",
+ "text": "CVF",
+ "bbox": [
+ 106,
+ 2,
+ 181,
+ 42
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "header",
+ "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.",
+ "bbox": [
+ 238,
+ 0,
+ 807,
+ 46
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "page_number",
+ "text": "14884",
+ "bbox": [
+ 480,
+ 944,
+ 519,
+ 955
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "[38]. Without knowing the quality of selected frames in advance, poses detected from them may be noisy. To tackle this issue, the current design employs a refine-recover-refine pipeline to obtain the final pose sequence.",
+ "bbox": [
+ 76,
+ 90,
+ 467,
+ 151
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Is there a way to simplify the pipeline and further reduce the computational cost? The two refinement modules for keyframe poses and recovered poses appear repetitive and unnecessary, as the poses in keyframes are refined twice. To address this, we propose a simplified recover-refine pipeline, as shown in Figure 1c. Within our framework, keyframes are selected using a sampler, and the pose sequence is recovered from detected poses in keyframes using an interpolator. The rough sequence is then refined using a single refinement module.",
+ "bbox": [
+ 75,
+ 152,
+ 467,
+ 303
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "We present a novel framework for sampling-based pose estimation utilizing a lightweight transformer encoder-like structure. The scaled dot product self-attention mechanism in the transformer network [34] is used to determine the relative importance of a token with respect to other tokens. However, the computational cost can be prohibitively high when using a large embedding size. To mitigate this problem, we replace the standard key-query attention with a synthetic self-attention module that generates attention weights using linear layers. To improve the quality of refinements, we synthesize both spatial and temporal attention matrices, which dynamically capture the inter-joint and inter-frame relationships. The features from both dimensions are combined and passed forward for further processing. We name our model MixSynthFormer. The main contributions of this work are summarized as follows:",
+ "bbox": [
+ 75,
+ 306,
+ 467,
+ 546
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "$\\triangleright$ We propose a highly-efficient keyframe-based pose estimation model MixSynthFormer following the transformer encoder structure. It can effectively estimate poses from sampled keyframe poses in a recover-refine pipeline.",
+ "bbox": [
+ 76,
+ 549,
+ 467,
+ 609
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "$\\triangleright$ We design an MLP-based mixed synthetic attention matrix generation module, MixSynth Attention, which generates attention matrices spatially and temporally from input representations. This design enables our model to dynamically fuse channel-wise and token-wise features and refine poses effectively. To speed up the computation, we introduce a reduction factor in the attention matrix generation, which further saves computation without compromising performance.",
+ "bbox": [
+ 75,
+ 611,
+ 467,
+ 746
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "$\\triangleright$ We validate our model on 2D and 3D pose estimation and body recovery tasks with four datasets and five estimators. Experimental results demonstrate that our model outperforms all keyframe-based pose estimation frameworks in terms of accuracy and efficiency. Additionally, our model can adapt to other motion synthesis tasks by indicating different keyframes. We conduct experiments on short-term motion prediction as a sub-task, and results show that it is also competitive compared to state-of-the-art motion prediction models.",
+ "bbox": [
+ 75,
+ 750,
+ 467,
+ 898
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2. Related Work",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 89,
+ 640,
+ 104
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2.1. Human Pose Estimation",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 114,
+ 720,
+ 130
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Pose estimation models can be categorized into two types based on how frames are fed into the model: frame-by-frame and keyframe-based estimators. Common pose estimation frameworks estimate poses in every frame and optimize efficiency through knowledge distillation [22, 27] or model structure design [3, 41, 29, 4, 37]. However, the computational costs are still high due to the frame-by-frame estimation. Besides, these methods are sensitive to occlusion cases in which partial joints are invisible in the camera view due to complex poses or interactions with environments, resulting in incoherent poses regardless of incorporating temporal information. In contrast, keyframe-based pose estimation frameworks exploit human motion continuity and can produce smoother sequences. Only keyframes are used for heavy single-frame estimation, while poses in the rest of the frames are recovered by a lightweight module. Thus, they can boost the efficiency of single-frame pose estimators in different tasks, such as SimplePose [36] for 2D poses, FCN [25] for 3D poses, and SPIN [20], PARE [19] and EFT [17] for body recovery.",
+ "bbox": [
+ 496,
+ 138,
+ 890,
+ 441
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Keyframe-based methods can be further divided into two categories depending on how the keyframes are selected: feature-based and sampling-based. Feature-based methods [40, 7] select \"good\" keyframes relying on intermediate feature representations and recover the whole sequence without refinement. KFPN [40] selects frames that capture global context based on image features extracted from inputs. MAPN [7] uses the internal motion signals and residual errors to determine whether the heavy pose estimation needs to be conducted on a frame. Nevertheless, the cost of feature extraction cannot be ignored.",
+ "bbox": [
+ 496,
+ 441,
+ 890,
+ 607
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Sampling-based methods [38] select keyframes using a uniform, random or customized sampling strategy. As selected frame may not be \"good\", these frameworks add a refinement module to clean noisy poses. Deciwatch [38] uses a standard transformer encoder to refine the uniformly sampled poses and a decoder to refine the recovered pose sequence based on denoised poses. The refinement on sampled keyframe poses is redundant and can be merged with the subsequent refinement module in the decoder. Moreover, Deciwatch [38] focuses on the temporal aspects and does not take spatial dependencies into consideration, which reflect interdependence among joints. Understanding the patterns of how joints are moving together can facilitate the refinement of incorrectly estimated joints.",
+ "bbox": [
+ 496,
+ 608,
+ 890,
+ 819
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2.2. Motion Completion and Prediction",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 830,
+ 803,
+ 847
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Motion completion aims to fill the motion of missing frames with specified keyframe constraints, presenting both past and future frames. Traditional methods utilize lin",
+ "bbox": [
+ 496,
+ 854,
+ 890,
+ 900
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "14885",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/901c1b48819d0e6a06a8ed5090a62be210b9b844d091b6db34c3ba83b1857d9f.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 106,
+ 93,
+ 421,
+ 351
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "image",
+ "img_path": "images/12957400a5b7ec853386d7cdcce8ef0746b514da3291f8c96551761a067b5473.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 437,
+ 95,
+ 633,
+ 333
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "image",
+ "img_path": "images/3c12733e75330dca013e530acd8e31ab4bd719562922a430ec766c729fb258f9.jpg",
+ "image_caption": [
+ "Figure 2: Overview of MixSynthFormer. MixSynthFormer recovers and refines poses from sparsely sampled keyframes. It consists of $L$ blocks of MixSynthEncoder blocks and two linear projection layers. MixSynthEncoder follows a standard transformer encoder structure but replaces multi-head self-attention with MixSynth Attention. MixSynth Attention synthesizes spatial and temporal attention via SynthAttention operations (depicted in Figure 3). The output from the two branches are combined and forwarded to a linear layer for feature fusion. (Detected poses in keyframes are colored, and recovered poses by interpolators are in gray.)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 651,
+ 95,
+ 867,
+ 354
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "ear, cubic-spline, Lagrange, and low-rank matrix completion, producing smooth transitions but lacking details [13]. Early learning-based methods employ RNN [9, 10] and convolutional models [12, 18]. Transformer-based models [6, 28, 30] are also adopted due to their exceptional performance in sequence-to-sequence problems. The aforementioned motion completion models are prevalent in the animation field and utilize ground-truth joint rotations and positions. However, in our case, the inputs are detected poses that can be noisy and aperiodic.",
+ "bbox": [
+ 75,
+ 440,
+ 468,
+ 590
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Motion prediction generates future motion given only historical frames. RNN [8, 24] and transformer-based [2, 1] methods are proposed to capture temporal dynamics. GCN-based methods, such as STSGCN [31] and STGAGCN [42], further model the spatial relationships among joints, obtaining high performance by closely combining spatio-temporal modeling.",
+ "bbox": [
+ 75,
+ 592,
+ 468,
+ 696
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Despite a range of frameworks, none of them treat motion completion and prediction as two mutually dependent tasks. Rather, each approach is specialized for one particular motion synthesis task. Notably, the sole distinction between motion completion and motion prediction pertains to the keyframe constraints. In this work, our model can handle both tasks with a unified structure.",
+ "bbox": [
+ 75,
+ 696,
+ 468,
+ 801
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "3. Method",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 814,
+ 168,
+ 829
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Problem Definition. Given an input video of $T$ frames, a sampler samples one frame out of every $K$ frames ( $T$ is a multiple of $K$ ). The sampled frames are regarded as keyframes where poses will be estimated by a pose esti",
+ "bbox": [
+ 75,
+ 839,
+ 468,
+ 902
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "mator. The detected poses from sampled frames may be unreliable due to complex poses or occlusion. We denote poses detected from sampled frames as $\\hat{\\mathbf{X}}_{\\text{noisy}}^{\\text{sampled}} \\in \\mathbb{R}^{\\frac{T}{K} \\times P}$ which serves as the input to our framework. $P$ is the number of pose parameters, which can denote joint positions (2D/3D) or rotations (6D) and may vary depending on the dataset and body representation. The objective of sampling-based pose estimation is to recover all the poses $\\hat{\\mathbf{X}} \\in \\mathbb{R}^{T \\times P}$ in the original sequence from detected poses $\\hat{\\mathbf{X}}_{\\text{noisy}}^{\\text{sampled}}$ in the sampled frames (referred to as sampled detected poses).",
+ "bbox": [
+ 496,
+ 440,
+ 890,
+ 592
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Model Overview. Figure 2 provides an overview of our proposed model. The goal of MixSynthFormer is to recover and refine the poses to obtain accurate estimations efficiently. Initially, the entire pose sequence is recovered through linear interpolation or a linear layer from the sampled detected poses. Those recovered poses are transformed into pose embeddings using linear projection and are fed into $L$ blocks of MixSynthEncoder. Each MixSynthEncoder comprises an attention block and a feed-forward network with residual connection, followed by a layer normalization layer. The attention block MixSynth Attention synthesizes spatial and temporal attention matrices concurrently. By integrating the spatial and temporal information, our model can refine the entire pose sequence in a simultaneous manner. Finally, the recovered poses are output via another linear layer.",
+ "bbox": [
+ 496,
+ 595,
+ 892,
+ 837
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "We next present details of MixSynthFormer, starting with an overview of the structure and then proceeding to the design of MixSynth Attention and the detailed operations within it. Finally, we describe the settings of using our",
+ "bbox": [
+ 496,
+ 839,
+ 893,
+ 900
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "14886",
+ "bbox": [
+ 480,
+ 944,
+ 519,
+ 955
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "model on motion prediction tasks, showcasing its adaptability to other motion synthesis tasks.",
+ "bbox": [
+ 76,
+ 90,
+ 468,
+ 121
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "3.1. Design of MixSynthFormer",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 130,
+ 323,
+ 146
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "We adopt a recover-refine pipeline in our framework. The whole pose sequence is preliminarily recovered along the temporal dimension by a traditional or a learned interpolator based on the detected poses $\\hat{\\mathbf{X}}_{\\text{noise}}^{\\text{sample}}$ in keyframes, denoted as $\\hat{\\mathbf{X}}_{\\text{noise}}^{\\text{recover}} \\in \\mathbb{R}^{T \\times P}$ :",
+ "bbox": [
+ 76,
+ 152,
+ 467,
+ 234
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {\\mathbf {X}} _ {n o i s y} ^ {r e c o v e r} = I n t e r p (\\hat {\\mathbf {X}} _ {n o i s y} ^ {s a m p l e d}) \\tag {1}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 166,
+ 243,
+ 468,
+ 266
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Recovering poses from sparse and noisy inputs results in inaccurate estimates, which necessitates additional processing to obtain a clean output. Moreover, traditional interpolation methods tend to produce overly smooth motion sequences that lack details. Hence, a refinement step is necessary to generate a more precise pose sequence, which can be formulated as a sequence-to-sequence problem. Transformers have shown remarkable success in solving such problems owing to their ability to capture global correlations among tokens. We design MixSynthFormer following the workflow of a standard transformer encoder, except for replacing the self-attention module with MixSynth Attention. Following are the high-level operations of MixSynthFormer.",
+ "bbox": [
+ 75,
+ 273,
+ 468,
+ 468
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "All poses in $\\hat{\\mathbf{X}}_{\\text {noisy }}^{\\text {recover }}$ are transformed to the pose embeddings $\\mathbf{Z}_0 \\in \\mathbb{R}^{T \\times C}$ through a linear layer with weights $\\mathbf{W}_0 \\in \\mathbb{R}^{P \\times C}$ , where $C$ is the dimension of the embedding space. Attention layers in transformers need position embeddings to specify the locations of input tokens, whereas our MLP-based MixSynth Attention is sensitive to the order of tokens in temporal attention synthesis like [33] and does not require additional position embeddings.",
+ "bbox": [
+ 75,
+ 469,
+ 468,
+ 592
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {Z} _ {0} = \\hat {\\mathbf {X}} _ {\\text {n o i s y}} ^ {\\text {r e c o v e r}} \\mathbf {W} _ {0} \\tag {2}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 205,
+ 604,
+ 468,
+ 625
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The pose embeddings are then passed to $L$ blocks of MixSynthEncoder. The model conforms to the conventional transformer encoder design, consisting of stacked MixSynth Attention module and feed-forward network (FFN) which is a two-layer point-wise fully connected layer with GELU activation [11]. Residual connection and layer normalization $(LN)$ are applied sequentially after every block.",
+ "bbox": [
+ 75,
+ 628,
+ 468,
+ 734
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {\\mathbf {Z}} _ {l} = L N \\left(M i x S y n t h A t t \\left(\\mathbf {Z} _ {l - 1}\\right) + \\mathbf {Z} _ {l - 1}\\right) \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 130,
+ 744,
+ 468,
+ 762
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {Z} _ {l} = L N \\left(F F N \\left(\\hat {\\mathbf {Z}} _ {l}\\right) + \\hat {\\mathbf {Z}} _ {l}\\right) \\tag {4}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 178,
+ 785,
+ 468,
+ 804
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Here $l \\in [1, L]$ denotes the $l$ -th block. $\\mathbf{Z}_{l-1}$ is the output from the previous encoder block, $\\hat{\\mathbf{Z}}_l$ is the intermediate representation and $\\mathbf{Z}_l$ is the output of the current encoder block. $\\mathbf{Z}_l$ and $\\hat{\\mathbf{Z}}_l$ have the same dimension in $\\mathbb{R}^{T \\times C}$ .",
+ "bbox": [
+ 75,
+ 809,
+ 468,
+ 869
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "After $L$ blocks of operations, the features in the embedding space $\\mathbf{Z}_L$ are transformed back to the pose dimension",
+ "bbox": [
+ 76,
+ 869,
+ 468,
+ 901
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "via another linear layer with weights $\\mathbf{W}_p\\in \\mathbb{R}^{C\\times P}$ . The use of the residual connection can expedite model convergence and yield more stable results. The final recovered pose sequence $\\hat{\\mathbf{X}}$ is thereby the summation of the output from MixSynthEncoder blocks and the preliminarily recovered noisy poses $\\hat{\\mathbf{X}}_{noisy}^{recover}$ :",
+ "bbox": [
+ 496,
+ 90,
+ 890,
+ 184
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {\\mathbf {X}} = \\left(L N (\\mathbf {Z} _ {L})\\right) \\mathbf {W} _ {p} + \\hat {\\mathbf {X}} _ {n o i s y} ^ {r e c o v e r} \\tag {5}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 584,
+ 194,
+ 890,
+ 215
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "3.2. MixSynth Attention",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 223,
+ 689,
+ 239
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Transformers are effective in capturing global correlations among tokens, but they may not be sufficient for refining noisy sequences. Token-wise attention primarily focuses on the temporal aspect, while local attention that captures the correlation among joints is also essential, particularly for recovering complex motions. To learn both global and local features, we propose MixSynth Attention, which mixes spatial and temporal synthetic attention matrices. The weights in these matrices reflect the relative inter-joint and inter-frame importance. The spatial matrix reveals the correlation among joints. When performing some motion patterns, joints that move together are highly correlated. It is helpful in locating a wrongly-detected joint from its related joints with accurate detection. The temporal matrix determines the impact of a frame with respect to other frames in the sequence, signifying the extent to which a frame affects the remaining frames.",
+ "bbox": [
+ 496,
+ 247,
+ 890,
+ 502
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "We explain the MixSynth Attention operations in detail. For simplicity, we eliminate the block index $l$ in the following explanation. The synthetic attention operation (SynthAttenOP) operates on $\\mathbf{Z} \\in \\mathbb{R}^{T \\times C}$ and outputs feature $\\mathbf{Y}_i \\in \\mathbb{R}^{T \\times C}$ , which can be spatial or temporal, denoted as $\\mathbf{Y}_s$ and $\\mathbf{Y}_t$ , respectively.",
+ "bbox": [
+ 496,
+ 503,
+ 890,
+ 594
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {Y} _ {i} = \\operatorname {S y n t h A t t e n O P} _ {i} (\\mathbf {Z}), i \\in \\{s, t \\} \\tag {6}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 566,
+ 606,
+ 890,
+ 623
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The spatially and temporally attended features are generated simultaneously. Inspired by [39], we apply split attention to re-balance the contribution of the two branches. Different motions may have different emphases. Inter-joint relations are more important for certain motions, while others may be more influenced by inter-frame relations. The weights $\\mathbf{A}_y\\in \\mathbb{R}^{2\\times C}$ of $\\mathbf{Y}_s$ and $\\mathbf{Y}_t$ are calculated as",
+ "bbox": [
+ 496,
+ 633,
+ 890,
+ 741
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {A} _ {y} = \\operatorname {S o f t m a x} (F F N (G A P (\\mathbf {Y} _ {s} + \\mathbf {Y} _ {t}))) \\tag {7}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 549,
+ 751,
+ 890,
+ 768
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $GAP$ denotes the global average pooling function, $FFN$ contains two fully connected layers with ReLU activation [26] and Softmax is for normalization. The weighted sum $\\mathbf{Y}_{out} \\in \\mathbb{R}^{T \\times C}$ is the summation of the element-wise product of the concatenated representations from two branches with the re-balancing weights $\\mathbf{A}_y$ :",
+ "bbox": [
+ 496,
+ 779,
+ 890,
+ 871
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {Y} _ {\\text {o u t}} = \\operatorname {S u m} \\left(\\mathbf {A} _ {y} \\circ \\left[ \\mathbf {Y} _ {s}; \\mathbf {Y} _ {t} \\right]\\right) \\tag {8}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 594,
+ 883,
+ 890,
+ 902
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "14887",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "The output of the MixSynth Attention block is attained by linearly transforming $\\mathbf{Y}_{out}$ with weights $\\mathbf{W}_o\\in \\mathbb{R}^{C\\times C}$ :",
+ "bbox": [
+ 76,
+ 90,
+ 468,
+ 122
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nM i x S y n t h A t t (\\mathbf {Z}) = \\mathbf {Y} _ {o u t} \\mathbf {W} _ {o} \\tag {9}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 166,
+ 131,
+ 468,
+ 148
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.3. SynthAttention Operation",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 156,
+ 316,
+ 174
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The self-attention module in transformers utilizes the pairwise dot product to calculate attention weights. The computation of the query-key matrix multiplication expands quadratically as the embedding size increases. To alleviate the impact of embedding size on computational cost, we propose synthetic attention operation (SynthAttenOp), whose computation grows linearly with the increase of embedding size, as a substitute for the standard attention. We discuss this part in detail in supplementary materials. SynthAttenOp generates attention matrices using linear layers. Details of its operation are described below.",
+ "bbox": [
+ 75,
+ 180,
+ 468,
+ 345
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "SynthAttenOp operates spatially and temporally. In spatial attention synthesis, input and output are transposed. Here we use $\\mathbf{Z} \\in \\mathbb{R}^{N \\times D}$ to illustrate the general case, where $N$ and $D$ can be substituted with either $C$ or $T$ . The coarse synthetic attention matrix $\\mathbf{A}_s \\in \\mathbb{R}^{N \\times N}$ is generated by a linear layer with weights $\\mathbf{W}_z \\in \\mathbb{R}^{D \\times N}$ :",
+ "bbox": [
+ 75,
+ 345,
+ 468,
+ 436
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {A} _ {s} = \\mathbf {Z W} _ {z} \\tag {10}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 228,
+ 446,
+ 468,
+ 463
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "As different frames or joints have different contributions, we apply the Squeeze-and-Excitation layer (SELayer) [14] to regulate the influence. It can amplify critical channels while mitigating the effect of less important ones. The fine-grained attention matrix $\\mathbf{A}_{\\text {synth }} \\in \\mathbb{R}^{N \\times N}$ is obtained by",
+ "bbox": [
+ 76,
+ 473,
+ 468,
+ 550
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\operatorname {S E L a y e r} \\left(\\mathbf {A} _ {s}\\right) = \\mathbf {A} _ {s} \\circ \\sigma \\left(\\mathbf {W} _ {e} \\delta \\left(\\mathbf {W} _ {s} G A P \\left(\\mathbf {A} _ {s}\\right)\\right)\\right) \\tag {11}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 93,
+ 570,
+ 468,
+ 587
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {A} _ {\\text {s y n t h}} = \\operatorname {S o f t m a x} (\\operatorname {S E L a y e r} (\\mathbf {A} _ {s})) \\tag {12}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 143,
+ 611,
+ 468,
+ 628
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where GAP denotes global average pooling and $GAP(\\mathbf{A}) \\in \\mathbb{R}^N$ , $\\delta$ and $\\sigma$ are Sigmoid activation and ReLU functions [26]. The reduction ratio $(r_{se})$ is an empirically found hyperparameter that limits model complexity [14]. $\\mathbf{W}_s \\in \\mathbb{R}^{\\frac{N}{r_{se}} \\times N}$ is for dimension reduction, and $\\mathbf{W}_e \\in \\mathbb{R}^{N \\times \\frac{N}{r_{se}}}$ is for expanding to the original dimension. The output of SELayer is a re-scaled input after channelwise multiplication with calculated weights. Softmax is inherited from the standard transformer, which normalizes the attention weights.",
+ "bbox": [
+ 76,
+ 633,
+ 468,
+ 787
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "In alignment with the standard transformer encoder, the output is the product of the attention matrix and the value:",
+ "bbox": [
+ 76,
+ 789,
+ 468,
+ 818
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {Y} = \\mathbf {A} _ {\\text {s y n t h}} \\mathbf {V} \\tag {13}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 218,
+ 829,
+ 468,
+ 845
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathbf{V} = \\mathbf{Z}\\in \\mathbb{R}^{N\\times D}$ and $\\mathbf{Y}\\in \\mathbb{R}^{N\\times D}$ .",
+ "bbox": [
+ 76,
+ 853,
+ 359,
+ 869
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The method described above is applicable for cases when $N$ is relatively small. When $N$ is large, the computation",
+ "bbox": [
+ 76,
+ 869,
+ 468,
+ 900
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/8a97663b6eccb8f6b299245b3ad8871f839592586046d9a46967aedc9c04cc0d.jpg",
+ "image_caption": [
+ "Figure 3: The procedure of reduced SynthAttenOp"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 501,
+ 90,
+ 890,
+ 150
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "cost grows quadratically owing to the matrix multiplication in Eqn. 13. We apply a reduction factor $r$ to the last dimension $D$ , saving the computation in Eqn. 13 by $r^2$ times. We denote the reduced dimension as $d = \\lfloor N / r \\rfloor$ . Figure 3 shows the process of reduced attention matrix generation and feature fusion. The reduced coarse attention matrix $\\mathbf{A}_s^r \\in \\mathbb{R}^{N \\times d}$ and reduced value $\\mathbf{V}^r \\in \\mathbb{R}^{d \\times D}$ are obtained through",
+ "bbox": [
+ 496,
+ 188,
+ 890,
+ 309
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {A} _ {s} ^ {r} = \\mathbf {Z} \\mathbf {W} _ {z} ^ {r}, \\mathbf {V} ^ {r} = \\mathbf {W} _ {v} \\mathbf {Z} \\tag {14}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 607,
+ 324,
+ 890,
+ 340
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where weights for reduced attention matrix generation and for reducing the value are $\\mathbf{W}_z^r\\in \\mathbb{R}^{D\\times d}$ and $\\mathbf{W}_v\\in \\mathbb{R}^{d\\times N}$ , respectively.",
+ "bbox": [
+ 496,
+ 345,
+ 890,
+ 391
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.4. Efficiency Calculation",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 400,
+ 705,
+ 416
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The overall cost of the model per frame combines the cost of a pose estimator and the cost of MixSynthFormer:",
+ "bbox": [
+ 498,
+ 422,
+ 890,
+ 453
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nF L O P s = \\frac {1}{K} * f (E s t.) + f (M S F) \\tag {15}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 566,
+ 460,
+ 890,
+ 491
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "$f(\\cdot)$ evaluates the FLOPs of a model per frame. As one frame out of $K$ frames is sampled for estimation, the per-frame cost of the estimator is $\\frac{1}{K} * f(Ext.$ . MixSynthFormer operates on every frame, recovering and refining poses on the whole sequence, and its FLOPs per frame is $f(MSF)$ . In particular, $f(MSF)$ is at least $10^{4}$ times smaller than $f(Ext.$ and is negligible compared to the heavy pose estimator. With $10\\%$ sampling ratio ( $K = 10$ ), our framework can speed up the computation by ten times.",
+ "bbox": [
+ 496,
+ 500,
+ 890,
+ 635
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.5. Loss Function",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 643,
+ 643,
+ 657
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Following [38], we apply L1 loss to minimize the error between the finely recovered pose sequence $\\hat{\\mathbf{X}}$ and the ground-truth pose sequence $\\mathbf{X}$ :",
+ "bbox": [
+ 496,
+ 667,
+ 890,
+ 713
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} | \\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t} | \\tag {16}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 619,
+ 722,
+ 890,
+ 762
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.6. Motion Prediction Setting",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 771,
+ 735,
+ 787
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Motion prediction and motion completion are two subfields in conditional motion generation with the only difference lying in the keyframe constraints. We include short-term motion prediction as a sub-task and demonstrate the versatility of our model on this task.",
+ "bbox": [
+ 496,
+ 795,
+ 890,
+ 869
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The input of motion prediction to our model is a historical ground-truth motion sequence $\\mathbf{X}_{1:T_h} \\in \\mathbb{R}^{T_h \\times P}$ , where",
+ "bbox": [
+ 498,
+ 869,
+ 890,
+ 901
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "14888",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "$T_{h}$ is the number of consecutive historical frames and $P$ is the size of pose parameters (i.e., 3D positions of human body joints). The goal is to predict the future sequence $\\mathbf{X}_{T_h + 1:T_h + T_f} \\in \\mathbb{R}^{T_f \\times P}$ given $\\mathbf{X}_{1:T_h}$ , where $T_{f}$ is the number of frames to be predicted. Similar to the pose estimation settings, the future poses are coarsely generated by interpolation, and then MixSynthFormer refines and outputs the future poses.",
+ "bbox": [
+ 75,
+ 90,
+ 468,
+ 210
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "We adopt the L2-norm between the predicted sequence and the ground-truth sequence calculated only on future frames as the loss function, which is equivalent to the Mean Per Joint Position Error (MPJPE) in [31, 42]:",
+ "bbox": [
+ 75,
+ 212,
+ 468,
+ 272
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} _ {f} = \\frac {1}{T _ {f}} \\sum_ {t = T _ {h} + 1} ^ {T _ {h} + T _ {f}} | | \\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t} | | _ {2} \\tag {17}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 169,
+ 281,
+ 468,
+ 325
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "We will test the above model in experiments, which showcases our model's adaptability to other synthesis tasks.",
+ "bbox": [
+ 76,
+ 334,
+ 468,
+ 364
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4. Experiments",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 377,
+ 209,
+ 393
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Datasets. We evaluate MixSynthFormer on four datasets for three tasks. For 2D pose estimation (to locate the 2D coordinates of joints from images), we use the Sub-JHMDB dataset [16] that contains daily and sports actions. For 3D pose estimation (to extend 2D pose data with depth information prediction), we use the widely-used Human3.6M indoor dataset [15]. For 3D body recovery (to estimate human shape and joint angles), we use the in-the-wild dataset 3DPW which includes videos taken by a moving phone camera [35] and AIST++ [21] with diverse dancing actions shot indoors.",
+ "bbox": [
+ 75,
+ 402,
+ 468,
+ 566
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Evaluation Metrics. We adopt the same evaluation metrics as in [38]. For 2D pose estimation, Percentage of the Correct Keypoints (PCK) is used, and three thresholds $(20\\%, 10\\%$ and $5\\%$ of the bounding box size under the pixel level) are set. For 3D tasks, Mean Per Joint Position Error (MPJPE) is utilized to report distance error and the mean Acceleration error (Accel) is adopted to reflect the smoothness of the estimated pose sequence. Mean Floating Point Operations (FLOPs (G)) per frame are also reported to show the computation costs.",
+ "bbox": [
+ 75,
+ 568,
+ 468,
+ 717
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Implementation Details. We use the same single-frame pose estimators and sampling strategy as in [38]. The uniform sampling ratio is set to $10\\%$ ( $K = 10$ ) in all experiments. The presence of the first and last frames is vital in recovery. In practice, we sample $Q$ frames with index $K \\times q$ , $q \\in [0, \\dots, Q - 1]$ from the input of length $K \\times Q + 1$ and additionally include the last frame of the input sequence as keyframes. The number of keyframes $Q$ , the embedding dimension $C$ and the number of MixSynthEncoder blocks $L$ vary among datasets and tasks. We train the model on a single A100 GPU with testing performed on an 8-core CPU Apple M1 Pro chip to obtain the inference time. Due to",
+ "bbox": [
+ 75,
+ 719,
+ 468,
+ 900
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "page limit, other detailed training settings and qualitative results are provided in supplementary materials.",
+ "bbox": [
+ 498,
+ 90,
+ 890,
+ 121
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.1. Comparison on 2D Pose Estimation",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 132,
+ 807,
+ 148
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Existing keyframe-based pose estimation methods use SimplePose [36] as the pose estimator. We follow the same experiment settings as in [40, 7, 38]. Table 1 presents the detailed PCK@0.2 results of each joint, the average performance of all three metrics and computation costs.",
+ "bbox": [
+ 496,
+ 155,
+ 890,
+ 229
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Compared with feature-based methods, KFP [40] and MAPN [7], our method improves the accuracy by over $4.6\\%$ in PCK@0.2. The improvements on fast-moving joints, such as elbows, wrists and ankles, are significant with a maximum increase of $14.8\\%$ . Those joints often cause image blur, leading to wrong detection results. Our sampling-based method incorporates a refinement module and can mitigate the effects of incorrect detection.",
+ "bbox": [
+ 496,
+ 231,
+ 890,
+ 349
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Compared to the sampling-based model DeciWatch [38], our model features a lighter design with 0.4M FLOPs computation, yet achieving a superior accuracy of $99.3\\%$ under PCK@0.2. Notably, our model achieves almost $100\\%$ in predicting the head and torso positions, which can be attributed to the integration of spatial dependencies. While the attention mechanism in [38] only considers inter-frame relations, our method also leverages inter-joint relations, which is particularly useful for limb position estimation. The accuracy of ankle increases from $96.5\\%$ to $97.8\\%$ . By fusing information from highly-correlated joints, our model also outperforms DeciWatch in strict metrics PCK@0.1 by $1\\%$ and PCK@0.05 by $1.3\\%$ .",
+ "bbox": [
+ 496,
+ 352,
+ 890,
+ 547
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.2. Comparison on 3D Pose Estimation",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 559,
+ 807,
+ 574
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "We next compare our method on 3D keypoint estimation and body recovery tasks against single-frame estimators [25, 20, 19, 17] and the sampling-based model DeciWatch [38]. Results on localization error and acceleration error are in Table 2.",
+ "bbox": [
+ 496,
+ 582,
+ 890,
+ 656
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Compared with DeciWatch [38], our model is significantly smaller with no performance degradation, thanks to SynthAttention. Our model achieves a 1-2mm reduction in MPJPE on Human3.6M and 3DPW without compromising smoothness. Notably, the FLOPs of the model trained on 3DPW are 0.03M, which is more than ten times smaller than DeciWatch. The inference time is less than 0.1 ms per frame on CPU. There are two main reasons for this: first, we only do refinements once; second, the attention matrix in our model is generated by the cost-efficient SynthAttenOp.",
+ "bbox": [
+ 496,
+ 657,
+ 890,
+ 809
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "On AIST++, our model reduces the acceleration error by $33\\%$ compared to DeciWatch and $86\\%$ compared to the single-frame estimator SPIN. We attribute the acceleration reduction to the integration of spatial and temporal attention. AIST++ contains poses with fast movements and diverse dancing actions. Due to the complexity of these poses,",
+ "bbox": [
+ 496,
+ 810,
+ 890,
+ 900
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "14889",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/79dc8f94be3cd51022fdafa4648a2583e80b8a4370be98329578526036851c42.jpg",
+ "table_caption": [
+ "Table 1: Comparison with keyframe-based pose estimation methods on Sub-JHMDB dataset [16] for 2D pose estimation. R means ResNet [39]. Ratio represents the sampling ratio. Sho., Elb., Wri., Ank., Avg. stand for shoulder, elbow, wrist, ankle and average PCK, respectively."
+ ],
+ "table_footnote": [],
+ "table_body": "| Dataset | Label | Train/Val/Test | Scenario | Evaluation metric |
| GroupEmoW [13] | Neutral, Positive, Negative | 11,127/3,178/1,589 | Specific social events | Acc |
| GAF 3.0 [13, 15] | Neutral, Positive, Negative | 9,815/4,349/3011 | Specific social events | Acc |
| HECO [44] | Surprise, Excitement, Happiness, Peace, Disgust, Anger, Fear, and Sadness | 6,570/1,877/938 | In-the-wild images that do not limit scenes | mAP |
| GAF Cohesion [13, 15] | 0-3 | 9,815/4,349/3011 | Specific social events | MSE |
",
+ "bbox": [
+ 81,
+ 88,
+ 890,
+ 181
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/f76662dc063f1e59ff4afeb130958930b1def33f11024e80663eaf0826c31fe6.jpg",
+ "image_caption": [
+ "Figure 5. Model parameter and accuracy trade-off comparison on the GAF 3.0 dataset. Our proposed models are highlighted by the blue rectangular box."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 84,
+ 247,
+ 460,
+ 449
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "CSwin [7]. In order to prove the generality of our approach, we have added the proposed dual-pathway and CPA module to the CSwin transformer [7], which is denoted by DCAT-CSwin-L.",
+ "bbox": [
+ 75,
+ 522,
+ 468,
+ 580
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "A comparison of the proposed model with the state-of-the-art group affect approaches on GAF 3.0 and GroupEmoW datasets can be found in Figure 5 and Table 3, respectively. Since our proposed DCAT model is based on a vision transformer, we also compare it with recent vision transformer models, such as multiscale-based (CrossViT [5], Focal-B [45], RegionViT-B [4], TCFormaler-L [47]), CNN-transformer hybrid-based (Conformer-L [32]), and lightweight-based models (EvoLeViT [43], DynamicSwin-T [36], DVT [40]).",
+ "bbox": [
+ 75,
+ 583,
+ 468,
+ 733
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Comparing with state-of-the-art group affect methods. Our DCAT-ViT and DCAT-CSwin-L achieve the best results compared to other group affect recognition methods on the GAF 3.0 dataset. Specifically, our DCAT-ViT achieves $77.29\\%$ accuracy with only 26.70M parameters, which is only about $7.6\\%$ of TMM [9], while our DCAT-CSwin-L with a larger number of parameters (358M) achieves $79.20\\%$ accuracy. On the GroupEmoW dataset, our DCAT-CSwin-L also achieves the best performance compared to other methods. It is worth noting that our proposed method only needs the global image and corre",
+ "bbox": [
+ 75,
+ 734,
+ 470,
+ 901
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/1d4c9f6bf51b19d2543c60b4f394b106c0db3e0613c511624adf03d8c086b5a9.jpg",
+ "table_caption": [
+ "Table 2. Datasets details used in this paper (Acc = Accuracy, mAP = mean average precision, MSE = mean square error)."
+ ],
+ "table_footnote": [],
+ "table_body": "| Dataset | Label | Train/Val/Test | Scenario | Evaluation metric |
| GroupEmoW [13] | Neutral, Positive, Negative | 11,127/3,178/1,589 | Specific social events | Acc |
| GAF 3.0 [13, 15] | Neutral, Positive, Negative | 9,815/4,349/3011 | Specific social events | Acc |
| HECO [44] | Surprise, Excitement, Happiness, Peace, Disgust, Anger, Fear, and Sadness | 6,570/1,877/938 | In-the-wild images that do not limit scenes | mAP |
| GAF Cohesion [13, 15] | 0-3 | 9,815/4,349/3011 | Specific social events | MSE |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.127,
+ 0.192,
+ 0.842,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "Table 2. Datasets details used in this paper (Acc = Accuracy, mAP = mean average precision, MSE = mean square error)."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.086,
+ 0.248,
+ 0.462,
+ 0.45
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.076,
+ 0.455,
+ 0.471,
+ 0.497
+ ],
+ "angle": 0,
+ "content": "Figure 5. Model parameter and accuracy trade-off comparison on the GAF 3.0 dataset. Our proposed models are highlighted by the blue rectangular box."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.523,
+ 0.47,
+ 0.581
+ ],
+ "angle": 0,
+ "content": "CSwin [7]. In order to prove the generality of our approach, we have added the proposed dual-pathway and CPA module to the CSwin transformer [7], which is denoted by DCAT-CSwin-L."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.584,
+ 0.47,
+ 0.734
+ ],
+ "angle": 0,
+ "content": "A comparison of the proposed model with the state-of-the-art group affect approaches on GAF 3.0 and GroupEmoW datasets can be found in Figure 5 and Table 3, respectively. Since our proposed DCAT model is based on a vision transformer, we also compare it with recent vision transformer models, such as multiscale-based (CrossViT [5], Focal-B [45], RegionViT-B [4], TCFormaler-L [47]), CNN-transformer hybrid-based (Conformer-L [32]), and lightweight-based models (EvoLeViT [43], DynamicSwin-T [36], DVT [40])."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.735,
+ 0.471,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "Comparing with state-of-the-art group affect methods. Our DCAT-ViT and DCAT-CSwin-L achieve the best results compared to other group affect recognition methods on the GAF 3.0 dataset. Specifically, our DCAT-ViT achieves \\(77.29\\%\\) accuracy with only 26.70M parameters, which is only about \\(7.6\\%\\) of TMM [9], while our DCAT-CSwin-L with a larger number of parameters (358M) achieves \\(79.20\\%\\) accuracy. On the GroupEmoW dataset, our DCAT-CSwin-L also achieves the best performance compared to other methods. It is worth noting that our proposed method only needs the global image and corre"
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.503,
+ 0.241,
+ 0.907,
+ 0.679
+ ],
+ "angle": 0,
+ "content": "| Dataset | Label | Train/Val/Test | Scenario | Evaluation metric |
| GroupEmoW [13] | Neutral, Positive, Negative | 11,127/3,178/1,589 | Specific social events | Acc |
| GAF 3.0 [13, 15] | Neutral, Positive, Negative | 9,815/4,349/3011 | Specific social events | Acc |
| HECO [44] | Surprise, Excitement, Happiness, Peace, Disgust, Anger, Fear, and Sadness | 6,570/1,877/938 | In-the-wild images that do not limit scenes | mAP |
| GAF Cohesion [13, 15] | 0-3 | 9,815/4,349/3011 | Specific social events | MSE |
+
+
+Figure 5. Model parameter and accuracy trade-off comparison on the GAF 3.0 dataset. Our proposed models are highlighted by the blue rectangular box.
+
+CSwin [7]. In order to prove the generality of our approach, we have added the proposed dual-pathway and CPA module to the CSwin transformer [7], which is denoted by DCAT-CSwin-L.
+
+A comparison of the proposed model with the state-of-the-art group affect approaches on GAF 3.0 and GroupEmoW datasets can be found in Figure 5 and Table 3, respectively. Since our proposed DCAT model is based on a vision transformer, we also compare it with recent vision transformer models, such as multiscale-based (CrossViT [5], Focal-B [45], RegionViT-B [4], TCFormaler-L [47]), CNN-transformer hybrid-based (Conformer-L [32]), and lightweight-based models (EvoLeViT [43], DynamicSwin-T [36], DVT [40]).
+
+Comparing with state-of-the-art group affect methods. Our DCAT-ViT and DCAT-CSwin-L achieve the best results compared to other group affect recognition methods on the GAF 3.0 dataset. Specifically, our DCAT-ViT achieves $77.29\%$ accuracy with only 26.70M parameters, which is only about $7.6\%$ of TMM [9], while our DCAT-CSwin-L with a larger number of parameters (358M) achieves $79.20\%$ accuracy. On the GroupEmoW dataset, our DCAT-CSwin-L also achieves the best performance compared to other methods. It is worth noting that our proposed method only needs the global image and corre
+
+Table 2. Datasets details used in this paper (Acc = Accuracy, mAP = mean average precision, MSE = mean square error).
+
+| Dataset | Label | Train/Val/Test | Scenario | Evaluation metric |
| GroupEmoW [13] | Neutral, Positive, Negative | 11,127/3,178/1,589 | Specific social events | Acc |
| GAF 3.0 [13, 15] | Neutral, Positive, Negative | 9,815/4,349/3011 | Specific social events | Acc |
| HECO [44] | Surprise, Excitement, Happiness, Peace, Disgust, Anger, Fear, and Sadness | 6,570/1,877/938 | In-the-wild images that do not limit scenes | mAP |
| GAF Cohesion [13, 15] | 0-3 | 9,815/4,349/3011 | Specific social events | MSE |
",
+ "image_path": "35e7517c0475054e767fcdb2ca7695acaf06e88d03c8be317365d0b030012ec6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 52,
+ 196,
+ 282,
+ 356
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 52,
+ 196,
+ 282,
+ 356
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 196,
+ 282,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 196,
+ 282,
+ 356
+ ],
+ "type": "image",
+ "image_path": "f76662dc063f1e59ff4afeb130958930b1def33f11024e80663eaf0826c31fe6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 46,
+ 360,
+ 288,
+ 393
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 360,
+ 288,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 360,
+ 288,
+ 393
+ ],
+ "type": "text",
+ "content": "Figure 5. Model parameter and accuracy trade-off comparison on the GAF 3.0 dataset. Our proposed models are highlighted by the blue rectangular box."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 46,
+ 414,
+ 287,
+ 460
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 414,
+ 287,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 414,
+ 287,
+ 460
+ ],
+ "type": "text",
+ "content": "CSwin [7]. In order to prove the generality of our approach, we have added the proposed dual-pathway and CPA module to the CSwin transformer [7], which is denoted by DCAT-CSwin-L."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 46,
+ 462,
+ 287,
+ 581
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 462,
+ 287,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 462,
+ 287,
+ 581
+ ],
+ "type": "text",
+ "content": "A comparison of the proposed model with the state-of-the-art group affect approaches on GAF 3.0 and GroupEmoW datasets can be found in Figure 5 and Table 3, respectively. Since our proposed DCAT model is based on a vision transformer, we also compare it with recent vision transformer models, such as multiscale-based (CrossViT [5], Focal-B [45], RegionViT-B [4], TCFormaler-L [47]), CNN-transformer hybrid-based (Conformer-L [32]), and lightweight-based models (EvoLeViT [43], DynamicSwin-T [36], DVT [40])."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": "Comparing with state-of-the-art group affect methods. Our DCAT-ViT and DCAT-CSwin-L achieve the best results compared to other group affect recognition methods on the GAF 3.0 dataset. Specifically, our DCAT-ViT achieves "
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "77.29\\%"
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": " accuracy with only 26.70M parameters, which is only about "
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "7.6\\%"
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": " of TMM [9], while our DCAT-CSwin-L with a larger number of parameters (358M) achieves "
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "79.20\\%"
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": " accuracy. On the GroupEmoW dataset, our DCAT-CSwin-L also achieves the best performance compared to other methods. It is worth noting that our proposed method only needs the global image and corre"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 307,
+ 190,
+ 555,
+ 537
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 77,
+ 152,
+ 515,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 77,
+ 152,
+ 515,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 77,
+ 152,
+ 515,
+ 163
+ ],
+ "type": "text",
+ "content": "Table 2. Datasets details used in this paper (Acc = Accuracy, mAP = mean average precision, MSE = mean square error)."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 307,
+ 190,
+ 555,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 190,
+ 555,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 190,
+ 555,
+ 537
+ ],
+ "type": "table",
+ "html": "| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 51.8 | 56.2 | 58.1 | 59.0 | 69.5 | 78.4 | 55.2 | 58.1 | 74.0 | 94.6 | 62.3 | 59.1 | 65.1 | 49.5 | 52.4 | 62.9 |
| Pavlakos et al. [82] CVPR'18 | 1 | 48.5 | 54.4 | 54.4 | 52.0 | 59.4 | 65.3 | 49.9 | 52.9 | 65.8 | 71.1 | 56.6 | 52.9 | 60.9 | 44.7 | 47.8 | 56.2 |
| LCN [27] ICCV'19 | 1 | 46.8 | 52.3 | 44.7 | 50.4 | 52.9 | 68.9 | 49.6 | 46.4 | 60.2 | 78.9 | 51.2 | 50.0 | 54.8 | 40.4 | 43.3 | 52.7 |
| Xu et al. [114] CVPR'21 | 1 | 45.2 | 49.9 | 47.5 | 50.9 | 54.9 | 66.1 | 48.5 | 46.3 | 59.7 | 71.5 | 51.4 | 48.6 | 53.9 | 39.9 | 44.1 | 51.9 |
| VideoPose3D [83] CVPR'19 | 243 | 45.2 | 46.7 | 43.3 | 45.6 | 48.1 | 55.1 | 44.6 | 44.3 | 57.3 | 65.8 | 47.1 | 44.0 | 49.0 | 32.8 | 33.9 | 46.8 |
| Cai et al. [12] ICCV'19 | 7 | 44.6 | 47.4 | 45.6 | 48.8 | 50.8 | 59.0 | 47.2 | 43.9 | 57.9 | 61.9 | 49.7 | 46.6 | 51.3 | 37.1 | 39.4 | 48.8 |
| Yeh et al. [122] NeurIPS'19 | 243 | 44.8 | 46.1 | 43.3 | 46.4 | 49.0 | 55.2 | 44.6 | 44.0 | 58.3 | 62.7 | 47.1 | 43.9 | 48.6 | 32.7 | 33.3 | 46.7 |
| Liu et al. [63] CVPR'20 | 243 | 41.8 | 44.8 | 41.1 | 44.9 | 47.4 | 54.1 | 43.4 | 42.2 | 56.2 | 63.6 | 45.3 | 43.5 | 45.3 | 31.3 | 32.2 | 45.1 |
| * Cheng et al. [21] AAAI'20 | 128 | 36.2 | 38.1 | 42.7 | 35.9 | 38.2 | 45.7 | 36.8 | 42.0 | 45.9 | 51.3 | 41.8 | 41.5 | 43.8 | 33.1 | 28.6 | 40.1 |
| * UGCN [109] ECCV'20 | 96 | 38.2 | 41.0 | 45.9 | 39.7 | 41.4 | 51.4 | 41.6 | 41.4 | 52.0 | 57.4 | 41.8 | 44.4 | 41.6 | 33.1 | 30.0 | 42.6 |
| † PoseFormer [127] ICCV'21 | 81 | 41.5 | 44.8 | 39.8 | 42.5 | 46.5 | 51.6 | 42.1 | 42.0 | 53.3 | 60.7 | 45.5 | 43.3 | 46.1 | 31.8 | 32.2 | 44.3 |
| * Wehrbein et al. [112] ICCV'21 | 200 | 38.5 | 42.5 | 39.9 | 41.7 | 46.5 | 51.6 | 39.9 | 40.8 | 49.5 | 56.8 | 45.3 | 46.4 | 46.8 | 37.8 | 40.4 | 44.3 |
| † MHFormer [52] CVPR'22 | 351 | 39.2 | 43.1 | 40.1 | 40.9 | 44.9 | 51.2 | 40.6 | 41.3 | 53.5 | 60.3 | 43.7 | 41.1 | 43.8 | 29.8 | 30.6 | 43.0 |
| *† MixSTE [126] CVPR'22 | 243 | 36.7 | 39.0 | 36.5 | 39.4 | 40.2 | 44.9 | 39.8 | 36.9 | 47.9 | 54.8 | 39.6 | 37.8 | 39.3 | 29.7 | 30.6 | 39.8 |
| † P-STMO [87] ECCV'22 | 243 | 38.4 | 42.1 | 39.8 | 40.2 | 45.2 | 48.9 | 40.4 | 38.3 | 53.8 | 57.3 | 43.9 | 41.6 | 42.2 | 29.3 | 29.3 | 42.1 |
| † Ours (scratch) | 243 | 36.3 | 38.7 | 38.6 | 33.6 | 42.1 | 50.1 | 36.2 | 35.7 | 50.1 | 56.6 | 41.3 | 37.4 | 37.7 | 25.6 | 26.5 | 39.2 |
| † Ours (finetune) | 243 | 36.1 | 37.5 | 35.8 | 32.1 | 40.3 | 46.3 | 36.1 | 35.3 | 46.9 | 53.9 | 39.5 | 36.3 | 35.8 | 25.1 | 25.3 | 37.5 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 37.7 | 44.4 | 40.3 | 42.1 | 48.2 | 54.9 | 44.4 | 42.1 | 54.6 | 58.0 | 45.1 | 46.4 | 47.6 | 36.4 | 40.4 | 45.5 |
| LCN [27] ICCV'19 | 1 | 36.3 | 38.8 | 29.7 | 37.8 | 34.6 | 42.5 | 39.8 | 32.5 | 36.2 | 39.5 | 34.4 | 38.4 | 38.2 | 31.3 | 34.2 | 36.3 |
| Xu et al. [114] CVPR'21 | 1 | 35.8 | 38.1 | 31.0 | 35.3 | 35.8 | 43.2 | 37.3 | 31.7 | 38.4 | 45.5 | 35.4 | 36.7 | 36.8 | 27.9 | 30.7 | 35.8 |
| UGCN [109] ECCV'20 | 96 | 23.0 | 25.7 | 22.8 | 22.6 | 24.1 | 30.6 | 24.9 | 24.5 | 31.1 | 35.0 | 25.6 | 24.3 | 25.1 | 19.8 | 18.4 | 25.6 |
| † PoseFormer [127] ICCV'21 | 81 | 30.0 | 33.6 | 29.9 | 31.0 | 30.2 | 33.3 | 34.8 | 31.4 | 37.8 | 38.6 | 31.7 | 31.5 | 29.0 | 23.3 | 23.1 | 31.3 |
| † MHFormer [52] CVPR'22 | 351 | 27.7 | 32.1 | 29.1 | 28.9 | 30.0 | 33.9 | 33.0 | 31.2 | 37.0 | 39.3 | 30.0 | 31.0 | 29.4 | 22.2 | 23.0 | 30.5 |
| † MixSTE [126] CVPR'22 | 243 | 21.6 | 22.0 | 20.4 | 21.0 | 20.8 | 24.3 | 24.7 | 21.9 | 26.9 | 24.9 | 21.2 | 21.5 | 20.8 | 14.7 | 15.6 | 21.6 |
| † P-STMO [87] ECCV'22 | 243 | 28.5 | 30.1 | 28.6 | 27.9 | 29.8 | 33.2 | 31.3 | 27.8 | 36.0 | 37.4 | 29.7 | 29.5 | 28.1 | 21.0 | 21.0 | 29.3 |
| † Ours (scratch) | 243 | 16.7 | 19.9 | 17.1 | 16.5 | 17.4 | 18.8 | 19.3 | 20.5 | 24.0 | 22.1 | 18.6 | 16.8 | 16.7 | 10.8 | 11.5 | 17.8 |
| † Ours (finetune) | 243 | 15.9 | 17.3 | 16.9 | 14.6 | 16.8 | 18.6 | 18.6 | 18.4 | 22.0 | 21.8 | 17.3 | 16.9 | 16.1 | 10.5 | 11.4 | 16.9 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| VideoPose3D [83] CVPR'19 | 243 | 3.0 | 3.1 | 2.2 | 3.4 | 2.3 | 2.7 | 2.7 | 3.1 | 2.1 | 2.9 | 2.3 | 2.4 | 3.7 | 3.1 | 2.8 | 2.8 |
| † PoseFormer [127] ICCV'21 | 81 | 3.2 | 3.4 | 2.6 | 3.6 | 2.6 | 3.0 | 2.9 | 3.2 | 2.6 | 3.3 | 2.7 | 2.7 | 3.8 | 3.2 | 2.9 | 3.1 |
| *† MixSTE [126] CVPR'22 | 243 | 2.5 | 2.7 | 1.9 | 2.8 | 1.9 | 2.2 | 2.3 | 2.6 | 1.6 | 2.2 | 1.9 | 2.0 | 3.1 | 2.6 | 2.2 | 2.3 |
| † Ours (scratch) | 243 | 1.8 | 2.1 | 1.5 | 2.0 | 1.5 | 1.9 | 1.8 | 2.1 | 1.2 | 1.8 | 1.5 | 1.4 | 2.6 | 2.0 | 1.7 | 1.8 |
| † Ours (finetune) | 243 | 1.7 | 1.9 | 1.4 | 1.9 | 1.4 | 1.7 | 1.7 | 1.9 | 1.1 | 1.6 | 1.4 | 1.3 | 2.4 | 1.9 | 1.6 | 1.7 |
",
+ "bbox": [
+ 127,
+ 88,
+ 841,
+ 483
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Table 1. Quantitative comparison of 3D human pose estimation on Human3.6M. (Top) MPJPE (mm) using detected 2D pose sequences. (Middle) MPJPE (mm) using GT 2D pose sequences. (Bottom) MPJVE (mm) using detected 2D pose sequences. $T$ denotes the clip length used by the method. We select the best results reported by each work. \\* denotes using HRNet [96] for 2D detection. $\\dagger$ denotes implemented with a spatio-temporal Transformer design. The best and second-best results are highlighted in bold and underlined formats.",
+ "bbox": [
+ 75,
+ 488,
+ 893,
+ 544
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "previous methods including other Transformer-based designs with spatio-temporal modeling. It shows the effectiveness of the proposed DSTformer in terms of learning 3D geometric structures and temporal dynamics. To further evaluate the upper bound of the models' capability, we compare the performance when using 2D GT pose sequences as input, which gets rid of the influence of different 2D detectors. As shown in Table 1 (middle), our models significantly outperform all the previous approaches. Table 1 (bottom) shows that both of our models also surpass previous works in terms of MPJVE, implying better temporal coherence. We attribute the performance advantage of our scratch model to the proposed DSTformer design. We include more comparisons and analysis to demonstrate the advantage of DSTformer with regard to other spatio-temporal architectures in Section 4.6 and supplementary materials. Additionally, our method achieves lower errors with the proposed pretraining stage.",
+ "bbox": [
+ 75,
+ 569,
+ 472,
+ 828
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.4. Skeleton-based Action Recognition",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 844,
+ 382,
+ 861
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "We further explore the possibility to learn action semantics with the pretrained human motion representations. We",
+ "bbox": [
+ 76,
+ 869,
+ 470,
+ 902
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "use the human action dataset NTU-RGB+D [86] which contains 57K videos of 60 action classes, and we follow the data splits Cross-subject (X-Sub) and Cross-view (X-View). The dataset has an extended version, NTU-RGB+D-120 [60], which contains 114K videos of 120 action classes. We follow the suggested One-shot action recognition protocol on NTU-RGB+D-120. For both datasets, we use HRNet [96] to extract 2D skeletons following [30]. Similarly, we train a scratch model with random initialization for comparison. As Table 2 (left) shows, our methods are comparable or superior to the state-of-the-art approaches. Notably, the pretraining stage accounts for a large performance gain.",
+ "bbox": [
+ 496,
+ 569,
+ 893,
+ 752
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Additionally, we delve into the one-shot setting which holds significant practical importance. Real-world applications often require fine-grained action recognition in specific domains such as education, sports, and healthcare. Unfortunately, the action classes in these scenarios are not typically defined in public datasets. As a result, only limited annotations for these novel action classes are available, making accurate recognition a challenging task. As proposed in [60], we report the results on the evaluation set of 20 novel classes",
+ "bbox": [
+ 496,
+ 763,
+ 895,
+ 902
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "15090",
+ "bbox": [
+ 480,
+ 944,
+ 519,
+ 957
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/51ad21c29c82282f9640cb41cd17642f9342b7ed66d8b5df8fd0782945ae5003.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 51.8 | 56.2 | 58.1 | 59.0 | 69.5 | 78.4 | 55.2 | 58.1 | 74.0 | 94.6 | 62.3 | 59.1 | 65.1 | 49.5 | 52.4 | 62.9 |
| Pavlakos et al. [82] CVPR'18 | 1 | 48.5 | 54.4 | 54.4 | 52.0 | 59.4 | 65.3 | 49.9 | 52.9 | 65.8 | 71.1 | 56.6 | 52.9 | 60.9 | 44.7 | 47.8 | 56.2 |
| LCN [27] ICCV'19 | 1 | 46.8 | 52.3 | 44.7 | 50.4 | 52.9 | 68.9 | 49.6 | 46.4 | 60.2 | 78.9 | 51.2 | 50.0 | 54.8 | 40.4 | 43.3 | 52.7 |
| Xu et al. [114] CVPR'21 | 1 | 45.2 | 49.9 | 47.5 | 50.9 | 54.9 | 66.1 | 48.5 | 46.3 | 59.7 | 71.5 | 51.4 | 48.6 | 53.9 | 39.9 | 44.1 | 51.9 |
| VideoPose3D [83] CVPR'19 | 243 | 45.2 | 46.7 | 43.3 | 45.6 | 48.1 | 55.1 | 44.6 | 44.3 | 57.3 | 65.8 | 47.1 | 44.0 | 49.0 | 32.8 | 33.9 | 46.8 |
| Cai et al. [12] ICCV'19 | 7 | 44.6 | 47.4 | 45.6 | 48.8 | 50.8 | 59.0 | 47.2 | 43.9 | 57.9 | 61.9 | 49.7 | 46.6 | 51.3 | 37.1 | 39.4 | 48.8 |
| Yeh et al. [122] NeurIPS'19 | 243 | 44.8 | 46.1 | 43.3 | 46.4 | 49.0 | 55.2 | 44.6 | 44.0 | 58.3 | 62.7 | 47.1 | 43.9 | 48.6 | 32.7 | 33.3 | 46.7 |
| Liu et al. [63] CVPR'20 | 243 | 41.8 | 44.8 | 41.1 | 44.9 | 47.4 | 54.1 | 43.4 | 42.2 | 56.2 | 63.6 | 45.3 | 43.5 | 45.3 | 31.3 | 32.2 | 45.1 |
| * Cheng et al. [21] AAAI'20 | 128 | 36.2 | 38.1 | 42.7 | 35.9 | 38.2 | 45.7 | 36.8 | 42.0 | 45.9 | 51.3 | 41.8 | 41.5 | 43.8 | 33.1 | 28.6 | 40.1 |
| * UGCN [109] ECCV'20 | 96 | 38.2 | 41.0 | 45.9 | 39.7 | 41.4 | 51.4 | 41.6 | 41.4 | 52.0 | 57.4 | 41.8 | 44.4 | 41.6 | 33.1 | 30.0 | 42.6 |
| † PoseFormer [127] ICCV'21 | 81 | 41.5 | 44.8 | 39.8 | 42.5 | 46.5 | 51.6 | 42.1 | 42.0 | 53.3 | 60.7 | 45.5 | 43.3 | 46.1 | 31.8 | 32.2 | 44.3 |
| * Wehrbein et al. [112] ICCV'21 | 200 | 38.5 | 42.5 | 39.9 | 41.7 | 46.5 | 51.6 | 39.9 | 40.8 | 49.5 | 56.8 | 45.3 | 46.4 | 46.8 | 37.8 | 40.4 | 44.3 |
| † MHFormer [52] CVPR'22 | 351 | 39.2 | 43.1 | 40.1 | 40.9 | 44.9 | 51.2 | 40.6 | 41.3 | 53.5 | 60.3 | 43.7 | 41.1 | 43.8 | 29.8 | 30.6 | 43.0 |
| *† MixSTE [126] CVPR'22 | 243 | 36.7 | 39.0 | 36.5 | 39.4 | 40.2 | 44.9 | 39.8 | 36.9 | 47.9 | 54.8 | 39.6 | 37.8 | 39.3 | 29.7 | 30.6 | 39.8 |
| † P-STMO [87] ECCV'22 | 243 | 38.4 | 42.1 | 39.8 | 40.2 | 45.2 | 48.9 | 40.4 | 38.3 | 53.8 | 57.3 | 43.9 | 41.6 | 42.2 | 29.3 | 29.3 | 42.1 |
| † Ours (scratch) | 243 | 36.3 | 38.7 | 38.6 | 33.6 | 42.1 | 50.1 | 36.2 | 35.7 | 50.1 | 56.6 | 41.3 | 37.4 | 37.7 | 25.6 | 26.5 | 39.2 |
| † Ours (finetune) | 243 | 36.1 | 37.5 | 35.8 | 32.1 | 40.3 | 46.3 | 36.1 | 35.3 | 46.9 | 53.9 | 39.5 | 36.3 | 35.8 | 25.1 | 25.3 | 37.5 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 37.7 | 44.4 | 40.3 | 42.1 | 48.2 | 54.9 | 44.4 | 42.1 | 54.6 | 58.0 | 45.1 | 46.4 | 47.6 | 36.4 | 40.4 | 45.5 |
| LCN [27] ICCV'19 | 1 | 36.3 | 38.8 | 29.7 | 37.8 | 34.6 | 42.5 | 39.8 | 32.5 | 36.2 | 39.5 | 34.4 | 38.4 | 38.2 | 31.3 | 34.2 | 36.3 |
| Xu et al. [114] CVPR'21 | 1 | 35.8 | 38.1 | 31.0 | 35.3 | 35.8 | 43.2 | 37.3 | 31.7 | 38.4 | 45.5 | 35.4 | 36.7 | 36.8 | 27.9 | 30.7 | 35.8 |
| UGCN [109] ECCV'20 | 96 | 23.0 | 25.7 | 22.8 | 22.6 | 24.1 | 30.6 | 24.9 | 24.5 | 31.1 | 35.0 | 25.6 | 24.3 | 25.1 | 19.8 | 18.4 | 25.6 |
| † PoseFormer [127] ICCV'21 | 81 | 30.0 | 33.6 | 29.9 | 31.0 | 30.2 | 33.3 | 34.8 | 31.4 | 37.8 | 38.6 | 31.7 | 31.5 | 29.0 | 23.3 | 23.1 | 31.3 |
| † MHFormer [52] CVPR'22 | 351 | 27.7 | 32.1 | 29.1 | 28.9 | 30.0 | 33.9 | 33.0 | 31.2 | 37.0 | 39.3 | 30.0 | 31.0 | 29.4 | 22.2 | 23.0 | 30.5 |
| † MixSTE [126] CVPR'22 | 243 | 21.6 | 22.0 | 20.4 | 21.0 | 20.8 | 24.3 | 24.7 | 21.9 | 26.9 | 24.9 | 21.2 | 21.5 | 20.8 | 14.7 | 15.6 | 21.6 |
| † P-STMO [87] ECCV'22 | 243 | 28.5 | 30.1 | 28.6 | 27.9 | 29.8 | 33.2 | 31.3 | 27.8 | 36.0 | 37.4 | 29.7 | 29.5 | 28.1 | 21.0 | 21.0 | 29.3 |
| † Ours (scratch) | 243 | 16.7 | 19.9 | 17.1 | 16.5 | 17.4 | 18.8 | 19.3 | 20.5 | 24.0 | 22.1 | 18.6 | 16.8 | 16.7 | 10.8 | 11.5 | 17.8 |
| † Ours (finetune) | 243 | 15.9 | 17.3 | 16.9 | 14.6 | 16.8 | 18.6 | 18.6 | 18.4 | 22.0 | 21.8 | 17.3 | 16.9 | 16.1 | 10.5 | 11.4 | 16.9 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| VideoPose3D [83] CVPR'19 | 243 | 3.0 | 3.1 | 2.2 | 3.4 | 2.3 | 2.7 | 2.7 | 3.1 | 2.1 | 2.9 | 2.3 | 2.4 | 3.7 | 3.1 | 2.8 | 2.8 |
| † PoseFormer [127] ICCV'21 | 81 | 3.2 | 3.4 | 2.6 | 3.6 | 2.6 | 3.0 | 2.9 | 3.2 | 2.6 | 3.3 | 2.7 | 2.7 | 3.8 | 3.2 | 2.9 | 3.1 |
| *† MixSTE [126] CVPR'22 | 243 | 2.5 | 2.7 | 1.9 | 2.8 | 1.9 | 2.2 | 2.3 | 2.6 | 1.6 | 2.2 | 1.9 | 2.0 | 3.1 | 2.6 | 2.2 | 2.3 |
| † Ours (scratch) | 243 | 1.8 | 2.1 | 1.5 | 2.0 | 1.5 | 1.9 | 1.8 | 2.1 | 1.2 | 1.8 | 1.5 | 1.4 | 2.6 | 2.0 | 1.7 | 1.8 |
| † Ours (finetune) | 243 | 1.7 | 1.9 | 1.4 | 1.9 | 1.4 | 1.7 | 1.7 | 1.9 | 1.1 | 1.6 | 1.4 | 1.3 | 2.4 | 1.9 | 1.6 | 1.7 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.076,
+ 0.489,
+ 0.894,
+ 0.545
+ ],
+ "angle": 0,
+ "content": "Table 1. Quantitative comparison of 3D human pose estimation on Human3.6M. (Top) MPJPE (mm) using detected 2D pose sequences. (Middle) MPJPE (mm) using GT 2D pose sequences. (Bottom) MPJVE (mm) using detected 2D pose sequences. \\(T\\) denotes the clip length used by the method. We select the best results reported by each work. \\* denotes using HRNet [96] for 2D detection. \\(\\dagger\\) denotes implemented with a spatio-temporal Transformer design. The best and second-best results are highlighted in bold and underlined formats."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.57,
+ 0.473,
+ 0.829
+ ],
+ "angle": 0,
+ "content": "previous methods including other Transformer-based designs with spatio-temporal modeling. It shows the effectiveness of the proposed DSTformer in terms of learning 3D geometric structures and temporal dynamics. To further evaluate the upper bound of the models' capability, we compare the performance when using 2D GT pose sequences as input, which gets rid of the influence of different 2D detectors. As shown in Table 1 (middle), our models significantly outperform all the previous approaches. Table 1 (bottom) shows that both of our models also surpass previous works in terms of MPJVE, implying better temporal coherence. We attribute the performance advantage of our scratch model to the proposed DSTformer design. We include more comparisons and analysis to demonstrate the advantage of DSTformer with regard to other spatio-temporal architectures in Section 4.6 and supplementary materials. Additionally, our method achieves lower errors with the proposed pretraining stage."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.845,
+ 0.383,
+ 0.862
+ ],
+ "angle": 0,
+ "content": "4.4. Skeleton-based Action Recognition"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.871,
+ 0.472,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "We further explore the possibility to learn action semantics with the pretrained human motion representations. We"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.57,
+ 0.895,
+ 0.753
+ ],
+ "angle": 0,
+ "content": "use the human action dataset NTU-RGB+D [86] which contains 57K videos of 60 action classes, and we follow the data splits Cross-subject (X-Sub) and Cross-view (X-View). The dataset has an extended version, NTU-RGB+D-120 [60], which contains 114K videos of 120 action classes. We follow the suggested One-shot action recognition protocol on NTU-RGB+D-120. For both datasets, we use HRNet [96] to extract 2D skeletons following [30]. Similarly, we train a scratch model with random initialization for comparison. As Table 2 (left) shows, our methods are comparable or superior to the state-of-the-art approaches. Notably, the pretraining stage accounts for a large performance gain."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.765,
+ 0.897,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "Additionally, we delve into the one-shot setting which holds significant practical importance. Real-world applications often require fine-grained action recognition in specific domains such as education, sports, and healthcare. Unfortunately, the action classes in these scenarios are not typically defined in public datasets. As a result, only limited annotations for these novel action classes are available, making accurate recognition a challenging task. As proposed in [60], we report the results on the evaluation set of 20 novel classes"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.52,
+ 0.958
+ ],
+ "angle": 0,
+ "content": "15090"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.151,
+ 0.089,
+ 0.409,
+ 0.258
+ ],
+ "angle": 0,
+ "content": "| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 51.8 | 56.2 | 58.1 | 59.0 | 69.5 | 78.4 | 55.2 | 58.1 | 74.0 | 94.6 | 62.3 | 59.1 | 65.1 | 49.5 | 52.4 | 62.9 |
| Pavlakos et al. [82] CVPR'18 | 1 | 48.5 | 54.4 | 54.4 | 52.0 | 59.4 | 65.3 | 49.9 | 52.9 | 65.8 | 71.1 | 56.6 | 52.9 | 60.9 | 44.7 | 47.8 | 56.2 |
| LCN [27] ICCV'19 | 1 | 46.8 | 52.3 | 44.7 | 50.4 | 52.9 | 68.9 | 49.6 | 46.4 | 60.2 | 78.9 | 51.2 | 50.0 | 54.8 | 40.4 | 43.3 | 52.7 |
| Xu et al. [114] CVPR'21 | 1 | 45.2 | 49.9 | 47.5 | 50.9 | 54.9 | 66.1 | 48.5 | 46.3 | 59.7 | 71.5 | 51.4 | 48.6 | 53.9 | 39.9 | 44.1 | 51.9 |
| VideoPose3D [83] CVPR'19 | 243 | 45.2 | 46.7 | 43.3 | 45.6 | 48.1 | 55.1 | 44.6 | 44.3 | 57.3 | 65.8 | 47.1 | 44.0 | 49.0 | 32.8 | 33.9 | 46.8 |
| Cai et al. [12] ICCV'19 | 7 | 44.6 | 47.4 | 45.6 | 48.8 | 50.8 | 59.0 | 47.2 | 43.9 | 57.9 | 61.9 | 49.7 | 46.6 | 51.3 | 37.1 | 39.4 | 48.8 |
| Yeh et al. [122] NeurIPS'19 | 243 | 44.8 | 46.1 | 43.3 | 46.4 | 49.0 | 55.2 | 44.6 | 44.0 | 58.3 | 62.7 | 47.1 | 43.9 | 48.6 | 32.7 | 33.3 | 46.7 |
| Liu et al. [63] CVPR'20 | 243 | 41.8 | 44.8 | 41.1 | 44.9 | 47.4 | 54.1 | 43.4 | 42.2 | 56.2 | 63.6 | 45.3 | 43.5 | 45.3 | 31.3 | 32.2 | 45.1 |
| * Cheng et al. [21] AAAI'20 | 128 | 36.2 | 38.1 | 42.7 | 35.9 | 38.2 | 45.7 | 36.8 | 42.0 | 45.9 | 51.3 | 41.8 | 41.5 | 43.8 | 33.1 | 28.6 | 40.1 |
| * UGCN [109] ECCV'20 | 96 | 38.2 | 41.0 | 45.9 | 39.7 | 41.4 | 51.4 | 41.6 | 41.4 | 52.0 | 57.4 | 41.8 | 44.4 | 41.6 | 33.1 | 30.0 | 42.6 |
| † PoseFormer [127] ICCV'21 | 81 | 41.5 | 44.8 | 39.8 | 42.5 | 46.5 | 51.6 | 42.1 | 42.0 | 53.3 | 60.7 | 45.5 | 43.3 | 46.1 | 31.8 | 32.2 | 44.3 |
| * Wehrbein et al. [112] ICCV'21 | 200 | 38.5 | 42.5 | 39.9 | 41.7 | 46.5 | 51.6 | 39.9 | 40.8 | 49.5 | 56.8 | 45.3 | 46.4 | 46.8 | 37.8 | 40.4 | 44.3 |
| † MHFormer [52] CVPR'22 | 351 | 39.2 | 43.1 | 40.1 | 40.9 | 44.9 | 51.2 | 40.6 | 41.3 | 53.5 | 60.3 | 43.7 | 41.1 | 43.8 | 29.8 | 30.6 | 43.0 |
| *† MixSTE [126] CVPR'22 | 243 | 36.7 | 39.0 | 36.5 | 39.4 | 40.2 | 44.9 | 39.8 | 36.9 | 47.9 | 54.8 | 39.6 | 37.8 | 39.3 | 29.7 | 30.6 | 39.8 |
| † P-STMO [87] ECCV'22 | 243 | 38.4 | 42.1 | 39.8 | 40.2 | 45.2 | 48.9 | 40.4 | 38.3 | 53.8 | 57.3 | 43.9 | 41.6 | 42.2 | 29.3 | 29.3 | 42.1 |
| † Ours (scratch) | 243 | 36.3 | 38.7 | 38.6 | 33.6 | 42.1 | 50.1 | 36.2 | 35.7 | 50.1 | 56.6 | 41.3 | 37.4 | 37.7 | 25.6 | 26.5 | 39.2 |
| † Ours (finetune) | 243 | 36.1 | 37.5 | 35.8 | 32.1 | 40.3 | 46.3 | 36.1 | 35.3 | 46.9 | 53.9 | 39.5 | 36.3 | 35.8 | 25.1 | 25.3 | 37.5 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 37.7 | 44.4 | 40.3 | 42.1 | 48.2 | 54.9 | 44.4 | 42.1 | 54.6 | 58.0 | 45.1 | 46.4 | 47.6 | 36.4 | 40.4 | 45.5 |
| LCN [27] ICCV'19 | 1 | 36.3 | 38.8 | 29.7 | 37.8 | 34.6 | 42.5 | 39.8 | 32.5 | 36.2 | 39.5 | 34.4 | 38.4 | 38.2 | 31.3 | 34.2 | 36.3 |
| Xu et al. [114] CVPR'21 | 1 | 35.8 | 38.1 | 31.0 | 35.3 | 35.8 | 43.2 | 37.3 | 31.7 | 38.4 | 45.5 | 35.4 | 36.7 | 36.8 | 27.9 | 30.7 | 35.8 |
| UGCN [109] ECCV'20 | 96 | 23.0 | 25.7 | 22.8 | 22.6 | 24.1 | 30.6 | 24.9 | 24.5 | 31.1 | 35.0 | 25.6 | 24.3 | 25.1 | 19.8 | 18.4 | 25.6 |
| † PoseFormer [127] ICCV'21 | 81 | 30.0 | 33.6 | 29.9 | 31.0 | 30.2 | 33.3 | 34.8 | 31.4 | 37.8 | 38.6 | 31.7 | 31.5 | 29.0 | 23.3 | 23.1 | 31.3 |
| † MHFormer [52] CVPR'22 | 351 | 27.7 | 32.1 | 29.1 | 28.9 | 30.0 | 33.9 | 33.0 | 31.2 | 37.0 | 39.3 | 30.0 | 31.0 | 29.4 | 22.2 | 23.0 | 30.5 |
| † MixSTE [126] CVPR'22 | 243 | 21.6 | 22.0 | 20.4 | 21.0 | 20.8 | 24.3 | 24.7 | 21.9 | 26.9 | 24.9 | 21.2 | 21.5 | 20.8 | 14.7 | 15.6 | 21.6 |
| † P-STMO [87] ECCV'22 | 243 | 28.5 | 30.1 | 28.6 | 27.9 | 29.8 | 33.2 | 31.3 | 27.8 | 36.0 | 37.4 | 29.7 | 29.5 | 28.1 | 21.0 | 21.0 | 29.3 |
| † Ours (scratch) | 243 | 16.7 | 19.9 | 17.1 | 16.5 | 17.4 | 18.8 | 19.3 | 20.5 | 24.0 | 22.1 | 18.6 | 16.8 | 16.7 | 10.8 | 11.5 | 17.8 |
| † Ours (finetune) | 243 | 15.9 | 17.3 | 16.9 | 14.6 | 16.8 | 18.6 | 18.6 | 18.4 | 22.0 | 21.8 | 17.3 | 16.9 | 16.1 | 10.5 | 11.4 | 16.9 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| VideoPose3D [83] CVPR'19 | 243 | 3.0 | 3.1 | 2.2 | 3.4 | 2.3 | 2.7 | 2.7 | 3.1 | 2.1 | 2.9 | 2.3 | 2.4 | 3.7 | 3.1 | 2.8 | 2.8 |
| † PoseFormer [127] ICCV'21 | 81 | 3.2 | 3.4 | 2.6 | 3.6 | 2.6 | 3.0 | 2.9 | 3.2 | 2.6 | 3.3 | 2.7 | 2.7 | 3.8 | 3.2 | 2.9 | 3.1 |
| *† MixSTE [126] CVPR'22 | 243 | 2.5 | 2.7 | 1.9 | 2.8 | 1.9 | 2.2 | 2.3 | 2.6 | 1.6 | 2.2 | 1.9 | 2.0 | 3.1 | 2.6 | 2.2 | 2.3 |
| † Ours (scratch) | 243 | 1.8 | 2.1 | 1.5 | 2.0 | 1.5 | 1.9 | 1.8 | 2.1 | 1.2 | 1.8 | 1.5 | 1.4 | 2.6 | 2.0 | 1.7 | 1.8 |
| † Ours (finetune) | 243 | 1.7 | 1.9 | 1.4 | 1.9 | 1.4 | 1.7 | 1.7 | 1.9 | 1.1 | 1.6 | 1.4 | 1.3 | 2.4 | 1.9 | 1.6 | 1.7 |
+
+Table 1. Quantitative comparison of 3D human pose estimation on Human3.6M. (Top) MPJPE (mm) using detected 2D pose sequences. (Middle) MPJPE (mm) using GT 2D pose sequences. (Bottom) MPJVE (mm) using detected 2D pose sequences. $T$ denotes the clip length used by the method. We select the best results reported by each work. \* denotes using HRNet [96] for 2D detection. $\dagger$ denotes implemented with a spatio-temporal Transformer design. The best and second-best results are highlighted in bold and underlined formats.
+
+previous methods including other Transformer-based designs with spatio-temporal modeling. It shows the effectiveness of the proposed DSTformer in terms of learning 3D geometric structures and temporal dynamics. To further evaluate the upper bound of the models' capability, we compare the performance when using 2D GT pose sequences as input, which gets rid of the influence of different 2D detectors. As shown in Table 1 (middle), our models significantly outperform all the previous approaches. Table 1 (bottom) shows that both of our models also surpass previous works in terms of MPJVE, implying better temporal coherence. We attribute the performance advantage of our scratch model to the proposed DSTformer design. We include more comparisons and analysis to demonstrate the advantage of DSTformer with regard to other spatio-temporal architectures in Section 4.6 and supplementary materials. Additionally, our method achieves lower errors with the proposed pretraining stage.
+
+# 4.4. Skeleton-based Action Recognition
+
+We further explore the possibility to learn action semantics with the pretrained human motion representations. We
+
+use the human action dataset NTU-RGB+D [86] which contains 57K videos of 60 action classes, and we follow the data splits Cross-subject (X-Sub) and Cross-view (X-View). The dataset has an extended version, NTU-RGB+D-120 [60], which contains 114K videos of 120 action classes. We follow the suggested One-shot action recognition protocol on NTU-RGB+D-120. For both datasets, we use HRNet [96] to extract 2D skeletons following [30]. Similarly, we train a scratch model with random initialization for comparison. As Table 2 (left) shows, our methods are comparable or superior to the state-of-the-art approaches. Notably, the pretraining stage accounts for a large performance gain.
+
+Additionally, we delve into the one-shot setting which holds significant practical importance. Real-world applications often require fine-grained action recognition in specific domains such as education, sports, and healthcare. Unfortunately, the action classes in these scenarios are not typically defined in public datasets. As a result, only limited annotations for these novel action classes are available, making accurate recognition a challenging task. As proposed in [60], we report the results on the evaluation set of 20 novel classes
+
+| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 51.8 | 56.2 | 58.1 | 59.0 | 69.5 | 78.4 | 55.2 | 58.1 | 74.0 | 94.6 | 62.3 | 59.1 | 65.1 | 49.5 | 52.4 | 62.9 |
| Pavlakos et al. [82] CVPR'18 | 1 | 48.5 | 54.4 | 54.4 | 52.0 | 59.4 | 65.3 | 49.9 | 52.9 | 65.8 | 71.1 | 56.6 | 52.9 | 60.9 | 44.7 | 47.8 | 56.2 |
| LCN [27] ICCV'19 | 1 | 46.8 | 52.3 | 44.7 | 50.4 | 52.9 | 68.9 | 49.6 | 46.4 | 60.2 | 78.9 | 51.2 | 50.0 | 54.8 | 40.4 | 43.3 | 52.7 |
| Xu et al. [114] CVPR'21 | 1 | 45.2 | 49.9 | 47.5 | 50.9 | 54.9 | 66.1 | 48.5 | 46.3 | 59.7 | 71.5 | 51.4 | 48.6 | 53.9 | 39.9 | 44.1 | 51.9 |
| VideoPose3D [83] CVPR'19 | 243 | 45.2 | 46.7 | 43.3 | 45.6 | 48.1 | 55.1 | 44.6 | 44.3 | 57.3 | 65.8 | 47.1 | 44.0 | 49.0 | 32.8 | 33.9 | 46.8 |
| Cai et al. [12] ICCV'19 | 7 | 44.6 | 47.4 | 45.6 | 48.8 | 50.8 | 59.0 | 47.2 | 43.9 | 57.9 | 61.9 | 49.7 | 46.6 | 51.3 | 37.1 | 39.4 | 48.8 |
| Yeh et al. [122] NeurIPS'19 | 243 | 44.8 | 46.1 | 43.3 | 46.4 | 49.0 | 55.2 | 44.6 | 44.0 | 58.3 | 62.7 | 47.1 | 43.9 | 48.6 | 32.7 | 33.3 | 46.7 |
| Liu et al. [63] CVPR'20 | 243 | 41.8 | 44.8 | 41.1 | 44.9 | 47.4 | 54.1 | 43.4 | 42.2 | 56.2 | 63.6 | 45.3 | 43.5 | 45.3 | 31.3 | 32.2 | 45.1 |
| * Cheng et al. [21] AAAI'20 | 128 | 36.2 | 38.1 | 42.7 | 35.9 | 38.2 | 45.7 | 36.8 | 42.0 | 45.9 | 51.3 | 41.8 | 41.5 | 43.8 | 33.1 | 28.6 | 40.1 |
| * UGCN [109] ECCV'20 | 96 | 38.2 | 41.0 | 45.9 | 39.7 | 41.4 | 51.4 | 41.6 | 41.4 | 52.0 | 57.4 | 41.8 | 44.4 | 41.6 | 33.1 | 30.0 | 42.6 |
| † PoseFormer [127] ICCV'21 | 81 | 41.5 | 44.8 | 39.8 | 42.5 | 46.5 | 51.6 | 42.1 | 42.0 | 53.3 | 60.7 | 45.5 | 43.3 | 46.1 | 31.8 | 32.2 | 44.3 |
| * Wehrbein et al. [112] ICCV'21 | 200 | 38.5 | 42.5 | 39.9 | 41.7 | 46.5 | 51.6 | 39.9 | 40.8 | 49.5 | 56.8 | 45.3 | 46.4 | 46.8 | 37.8 | 40.4 | 44.3 |
| † MHFormer [52] CVPR'22 | 351 | 39.2 | 43.1 | 40.1 | 40.9 | 44.9 | 51.2 | 40.6 | 41.3 | 53.5 | 60.3 | 43.7 | 41.1 | 43.8 | 29.8 | 30.6 | 43.0 |
| *† MixSTE [126] CVPR'22 | 243 | 36.7 | 39.0 | 36.5 | 39.4 | 40.2 | 44.9 | 39.8 | 36.9 | 47.9 | 54.8 | 39.6 | 37.8 | 39.3 | 29.7 | 30.6 | 39.8 |
| † P-STMO [87] ECCV'22 | 243 | 38.4 | 42.1 | 39.8 | 40.2 | 45.2 | 48.9 | 40.4 | 38.3 | 53.8 | 57.3 | 43.9 | 41.6 | 42.2 | 29.3 | 29.3 | 42.1 |
| † Ours (scratch) | 243 | 36.3 | 38.7 | 38.6 | 33.6 | 42.1 | 50.1 | 36.2 | 35.7 | 50.1 | 56.6 | 41.3 | 37.4 | 37.7 | 25.6 | 26.5 | 39.2 |
| † Ours (finetune) | 243 | 36.1 | 37.5 | 35.8 | 32.1 | 40.3 | 46.3 | 36.1 | 35.3 | 46.9 | 53.9 | 39.5 | 36.3 | 35.8 | 25.1 | 25.3 | 37.5 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| Martinez et al. [73] ICCV'17 | 1 | 37.7 | 44.4 | 40.3 | 42.1 | 48.2 | 54.9 | 44.4 | 42.1 | 54.6 | 58.0 | 45.1 | 46.4 | 47.6 | 36.4 | 40.4 | 45.5 |
| LCN [27] ICCV'19 | 1 | 36.3 | 38.8 | 29.7 | 37.8 | 34.6 | 42.5 | 39.8 | 32.5 | 36.2 | 39.5 | 34.4 | 38.4 | 38.2 | 31.3 | 34.2 | 36.3 |
| Xu et al. [114] CVPR'21 | 1 | 35.8 | 38.1 | 31.0 | 35.3 | 35.8 | 43.2 | 37.3 | 31.7 | 38.4 | 45.5 | 35.4 | 36.7 | 36.8 | 27.9 | 30.7 | 35.8 |
| UGCN [109] ECCV'20 | 96 | 23.0 | 25.7 | 22.8 | 22.6 | 24.1 | 30.6 | 24.9 | 24.5 | 31.1 | 35.0 | 25.6 | 24.3 | 25.1 | 19.8 | 18.4 | 25.6 |
| † PoseFormer [127] ICCV'21 | 81 | 30.0 | 33.6 | 29.9 | 31.0 | 30.2 | 33.3 | 34.8 | 31.4 | 37.8 | 38.6 | 31.7 | 31.5 | 29.0 | 23.3 | 23.1 | 31.3 |
| † MHFormer [52] CVPR'22 | 351 | 27.7 | 32.1 | 29.1 | 28.9 | 30.0 | 33.9 | 33.0 | 31.2 | 37.0 | 39.3 | 30.0 | 31.0 | 29.4 | 22.2 | 23.0 | 30.5 |
| † MixSTE [126] CVPR'22 | 243 | 21.6 | 22.0 | 20.4 | 21.0 | 20.8 | 24.3 | 24.7 | 21.9 | 26.9 | 24.9 | 21.2 | 21.5 | 20.8 | 14.7 | 15.6 | 21.6 |
| † P-STMO [87] ECCV'22 | 243 | 28.5 | 30.1 | 28.6 | 27.9 | 29.8 | 33.2 | 31.3 | 27.8 | 36.0 | 37.4 | 29.7 | 29.5 | 28.1 | 21.0 | 21.0 | 29.3 |
| † Ours (scratch) | 243 | 16.7 | 19.9 | 17.1 | 16.5 | 17.4 | 18.8 | 19.3 | 20.5 | 24.0 | 22.1 | 18.6 | 16.8 | 16.7 | 10.8 | 11.5 | 17.8 |
| † Ours (finetune) | 243 | 15.9 | 17.3 | 16.9 | 14.6 | 16.8 | 18.6 | 18.6 | 18.4 | 22.0 | 21.8 | 17.3 | 16.9 | 16.1 | 10.5 | 11.4 | 16.9 |
| Method | T | Dire. | Disc. | Eat | Greet | Phone | Photo | Pose | Purch. | Sit | SitD | Smoke | Wait | WalkD | Walk | WalkT | Avg |
| VideoPose3D [83] CVPR'19 | 243 | 3.0 | 3.1 | 2.2 | 3.4 | 2.3 | 2.7 | 2.7 | 3.1 | 2.1 | 2.9 | 2.3 | 2.4 | 3.7 | 3.1 | 2.8 | 2.8 |
| † PoseFormer [127] ICCV'21 | 81 | 3.2 | 3.4 | 2.6 | 3.6 | 2.6 | 3.0 | 2.9 | 3.2 | 2.6 | 3.3 | 2.7 | 2.7 | 3.8 | 3.2 | 2.9 | 3.1 |
| *† MixSTE [126] CVPR'22 | 243 | 2.5 | 2.7 | 1.9 | 2.8 | 1.9 | 2.2 | 2.3 | 2.6 | 1.6 | 2.2 | 1.9 | 2.0 | 3.1 | 2.6 | 2.2 | 2.3 |
| † Ours (scratch) | 243 | 1.8 | 2.1 | 1.5 | 2.0 | 1.5 | 1.9 | 1.8 | 2.1 | 1.2 | 1.8 | 1.5 | 1.4 | 2.6 | 2.0 | 1.7 | 1.8 |
| † Ours (finetune) | 243 | 1.7 | 1.9 | 1.4 | 1.9 | 1.4 | 1.7 | 1.7 | 1.9 | 1.1 | 1.6 | 1.4 | 1.3 | 2.4 | 1.9 | 1.6 | 1.7 |
",
+ "image_path": "04e0a27d94520649ec2e26314db2bd89480c373df24c9d50046184337ed66a79.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 46,
+ 387,
+ 547,
+ 431
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 387,
+ 547,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 387,
+ 547,
+ 431
+ ],
+ "type": "text",
+ "content": "Table 1. Quantitative comparison of 3D human pose estimation on Human3.6M. (Top) MPJPE (mm) using detected 2D pose sequences. (Middle) MPJPE (mm) using GT 2D pose sequences. (Bottom) MPJVE (mm) using detected 2D pose sequences. "
+ },
+ {
+ "bbox": [
+ 46,
+ 387,
+ 547,
+ 431
+ ],
+ "type": "inline_equation",
+ "content": "T"
+ },
+ {
+ "bbox": [
+ 46,
+ 387,
+ 547,
+ 431
+ ],
+ "type": "text",
+ "content": " denotes the clip length used by the method. We select the best results reported by each work. \\* denotes using HRNet [96] for 2D detection. "
+ },
+ {
+ "bbox": [
+ 46,
+ 387,
+ 547,
+ 431
+ ],
+ "type": "inline_equation",
+ "content": "\\dagger"
+ },
+ {
+ "bbox": [
+ 46,
+ 387,
+ 547,
+ 431
+ ],
+ "type": "text",
+ "content": " denotes implemented with a spatio-temporal Transformer design. The best and second-best results are highlighted in bold and underlined formats."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 46,
+ 451,
+ 289,
+ 656
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 451,
+ 289,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 451,
+ 289,
+ 656
+ ],
+ "type": "text",
+ "content": "previous methods including other Transformer-based designs with spatio-temporal modeling. It shows the effectiveness of the proposed DSTformer in terms of learning 3D geometric structures and temporal dynamics. To further evaluate the upper bound of the models' capability, we compare the performance when using 2D GT pose sequences as input, which gets rid of the influence of different 2D detectors. As shown in Table 1 (middle), our models significantly outperform all the previous approaches. Table 1 (bottom) shows that both of our models also surpass previous works in terms of MPJVE, implying better temporal coherence. We attribute the performance advantage of our scratch model to the proposed DSTformer design. We include more comparisons and analysis to demonstrate the advantage of DSTformer with regard to other spatio-temporal architectures in Section 4.6 and supplementary materials. Additionally, our method achieves lower errors with the proposed pretraining stage."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 47,
+ 669,
+ 234,
+ 682
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 669,
+ 234,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 669,
+ 234,
+ 682
+ ],
+ "type": "text",
+ "content": "4.4. Skeleton-based Action Recognition"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 47,
+ 689,
+ 288,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 689,
+ 288,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 689,
+ 288,
+ 715
+ ],
+ "type": "text",
+ "content": "We further explore the possibility to learn action semantics with the pretrained human motion representations. We"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 304,
+ 451,
+ 547,
+ 596
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 451,
+ 547,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 451,
+ 547,
+ 596
+ ],
+ "type": "text",
+ "content": "use the human action dataset NTU-RGB+D [86] which contains 57K videos of 60 action classes, and we follow the data splits Cross-subject (X-Sub) and Cross-view (X-View). The dataset has an extended version, NTU-RGB+D-120 [60], which contains 114K videos of 120 action classes. We follow the suggested One-shot action recognition protocol on NTU-RGB+D-120. For both datasets, we use HRNet [96] to extract 2D skeletons following [30]. Similarly, we train a scratch model with random initialization for comparison. As Table 2 (left) shows, our methods are comparable or superior to the state-of-the-art approaches. Notably, the pretraining stage accounts for a large performance gain."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 304,
+ 605,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 605,
+ 548,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 605,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": "Additionally, we delve into the one-shot setting which holds significant practical importance. Real-world applications often require fine-grained action recognition in specific domains such as education, sports, and healthcare. Unfortunately, the action classes in these scenarios are not typically defined in public datasets. As a result, only limited annotations for these novel action classes are available, making accurate recognition a challenging task. As proposed in [60], we report the results on the evaluation set of 20 novel classes"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 318,
+ 758
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 318,
+ 758
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 748,
+ 318,
+ 758
+ ],
+ "type": "text",
+ "content": "15090"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 92,
+ 70,
+ 250,
+ 204
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 92,
+ 70,
+ 250,
+ 204
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 92,
+ 70,
+ 250,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 92,
+ 70,
+ 250,
+ 204
+ ],
+ "type": "table",
+ "html": "| Variant | PSNR |
| Base + FFT | 32.35 |
| Base + FFT w/. Learnable filter | 32.48 |
| Base + FFT w/. Learnable filter + original feature | 32.49 |
| Base + FFT w/. Learnable filter + BN | 32.43 |
| Base + FFT w/. Learnable filter + IN | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 3 × 3 | 32.48 |
| Base + FFT w/. Learnable filter + IN + DWConv 7 × 7 | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 11 × 11 | 32.51 |
| Base + FFT w/. Learnable filter + IN + DWConv 21 × 21 (Our SFE) | 32.55 |
| Base + w/. Learnable filter + IN + DWConv 21 × 21 | 31.93 |
",
+ "bbox": [
+ 80,
+ 375,
+ 467,
+ 497
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "table",
+ "img_path": "images/352994fb6f1c80049cf246493577855da85ce7c18056f50f31d337722b244221.jpg",
+ "table_caption": [
+ "(b) Effects of the SFE branch."
+ ],
+ "table_footnote": [],
+ "table_body": "| Variant | PSNR |
| Base + Energy function | 32.40 |
| Base + DCT w/. Fixed coefficients | 32.32 |
| Base + DCT w/. Learnable filter | 32.41 |
| Base + DCT w/. Fixed coefficients + Energy function | 32.37 |
| Base + DCT w/. Learnable filter + Energy function | 32.50 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 3 × 3 | 32.42 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 7 × 7 | 32.40 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 11 × 11 (Our EFE) | 32.52 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 21 × 21 | 32.43 |
| Base + w/. Learnable filter + Energy function + DWConv 11 × 11 | 32.07 |
",
+ "bbox": [
+ 80,
+ 517,
+ 467,
+ 621
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "table",
+ "img_path": "images/80eed0450f7d8ae9423672e94e957d215863e09366ec6d8a06fe86cafe7b3ae2.jpg",
+ "table_caption": [
+ "(c) Effects of the EFE branch.",
+ "Table 3: Comparison between the proposed privilege training and previous training methods with BasicVSR on [38]."
+ ],
+ "table_footnote": [],
+ "table_body": "| Variant | PSNR |
| Base + FFT | 32.35 |
| Base + FFT w/. Learnable filter | 32.48 |
| Base + FFT w/. Learnable filter + original feature | 32.49 |
| Base + FFT w/. Learnable filter + BN | 32.43 |
| Base + FFT w/. Learnable filter + IN | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 3 × 3 | 32.48 |
| Base + FFT w/. Learnable filter + IN + DWConv 7 × 7 | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 11 × 11 | 32.51 |
| Base + FFT w/. Learnable filter + IN + DWConv 21 × 21 (Our SFE) | 32.55 |
| Base + w/. Learnable filter + IN + DWConv 21 × 21 | 31.93 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.185,
+ 0.499,
+ 0.362,
+ 0.51
+ ],
+ "angle": 0,
+ "content": "(b) Effects of the SFE branch."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.082,
+ 0.518,
+ 0.468,
+ 0.622
+ ],
+ "angle": 0,
+ "content": "| Variant | PSNR |
| Base + Energy function | 32.40 |
| Base + DCT w/. Fixed coefficients | 32.32 |
| Base + DCT w/. Learnable filter | 32.41 |
| Base + DCT w/. Fixed coefficients + Energy function | 32.37 |
| Base + DCT w/. Learnable filter + Energy function | 32.50 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 3 × 3 | 32.42 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 7 × 7 | 32.40 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 11 × 11 (Our EFE) | 32.52 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 21 × 21 | 32.43 |
| Base + w/. Learnable filter + Energy function + DWConv 11 × 11 | 32.07 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.184,
+ 0.623,
+ 0.362,
+ 0.633
+ ],
+ "angle": 0,
+ "content": "(c) Effects of the EFE branch."
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.077,
+ 0.657,
+ 0.469,
+ 0.686
+ ],
+ "angle": 0,
+ "content": "Table 3: Comparison between the proposed privilege training and previous training methods with BasicVSR on [38]."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.08,
+ 0.695,
+ 0.469,
+ 0.8
+ ],
+ "angle": 0,
+ "content": "| Variant | PSNR |
| Base + FFT | 32.35 |
| Base + FFT w/. Learnable filter | 32.48 |
| Base + FFT w/. Learnable filter + original feature | 32.49 |
| Base + FFT w/. Learnable filter + BN | 32.43 |
| Base + FFT w/. Learnable filter + IN | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 3 × 3 | 32.48 |
| Base + FFT w/. Learnable filter + IN + DWConv 7 × 7 | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 11 × 11 | 32.51 |
| Base + FFT w/. Learnable filter + IN + DWConv 21 × 21 (Our SFE) | 32.55 |
| Base + w/. Learnable filter + IN + DWConv 21 × 21 | 31.93 |
+
+(b) Effects of the SFE branch.
+
+| Variant | PSNR |
| Base + Energy function | 32.40 |
| Base + DCT w/. Fixed coefficients | 32.32 |
| Base + DCT w/. Learnable filter | 32.41 |
| Base + DCT w/. Fixed coefficients + Energy function | 32.37 |
| Base + DCT w/. Learnable filter + Energy function | 32.50 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 3 × 3 | 32.42 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 7 × 7 | 32.40 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 11 × 11 (Our EFE) | 32.52 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 21 × 21 | 32.43 |
| Base + w/. Learnable filter + Energy function + DWConv 11 × 11 | 32.07 |
+
+(c) Effects of the EFE branch.
+Table 3: Comparison between the proposed privilege training and previous training methods with BasicVSR on [38].
+
+| Variant | PSNR |
| Base + FFT | 32.35 |
| Base + FFT w/. Learnable filter | 32.48 |
| Base + FFT w/. Learnable filter + original feature | 32.49 |
| Base + FFT w/. Learnable filter + BN | 32.43 |
| Base + FFT w/. Learnable filter + IN | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 3 × 3 | 32.48 |
| Base + FFT w/. Learnable filter + IN + DWConv 7 × 7 | 32.52 |
| Base + FFT w/. Learnable filter + IN + DWConv 11 × 11 | 32.51 |
| Base + FFT w/. Learnable filter + IN + DWConv 21 × 21 (Our SFE) | 32.55 |
| Base + w/. Learnable filter + IN + DWConv 21 × 21 | 31.93 |
",
+ "image_path": "caa2452404a354f8cfa898d4543dc10bcd529400c67fa60e13ee294e724ec2d7.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 48,
+ 436,
+ 278,
+ 523
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 110,
+ 420,
+ 215,
+ 429
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 420,
+ 215,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 420,
+ 215,
+ 429
+ ],
+ "type": "text",
+ "content": "(b) Effects of the SFE branch."
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 48,
+ 436,
+ 278,
+ 523
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 436,
+ 278,
+ 523
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 436,
+ 278,
+ 523
+ ],
+ "type": "table",
+ "html": "| Variant | PSNR |
| Base + Energy function | 32.40 |
| Base + DCT w/. Fixed coefficients | 32.32 |
| Base + DCT w/. Learnable filter | 32.41 |
| Base + DCT w/. Fixed coefficients + Energy function | 32.37 |
| Base + DCT w/. Learnable filter + Energy function | 32.50 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 3 × 3 | 32.42 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 7 × 7 | 32.40 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 11 × 11 (Our EFE) | 32.52 |
| Base + DCT w/. Learnable filter + Energy function + DWConv 21 × 21 | 32.43 |
| Base + w/. Learnable filter + Energy function + DWConv 11 × 11 | 32.07 |
",
+ "image_path": "352994fb6f1c80049cf246493577855da85ce7c18056f50f31d337722b244221.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 47,
+ 585,
+ 279,
+ 673
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 109,
+ 524,
+ 215,
+ 532
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 524,
+ 215,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 524,
+ 215,
+ 532
+ ],
+ "type": "text",
+ "content": "(c) Effects of the EFE branch."
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 45,
+ 553,
+ 279,
+ 577
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 553,
+ 279,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 553,
+ 279,
+ 577
+ ],
+ "type": "text",
+ "content": "Table 3: Comparison between the proposed privilege training and previous training methods with BasicVSR on [38]."
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 47,
+ 585,
+ 279,
+ 673
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 585,
+ 279,
+ 673
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 585,
+ 279,
+ 673
+ ],
+ "type": "table",
+ "html": "| Modality | Task | Object | Scenario |
| Object category | Object selection | Multiple objects | Pixel correspondence | DoF | Occlusion |
| RGB | MOT [83] | Limited, usually pedestrians or vehicles | Detected | Y | N | 4 | Y |
| MOTS [75] | Limited, usually pedestrians or vehicles | Detected | Y | N | N/A | N |
| VOS [66] | Arbitrary | User-specified in the first frame | N | N | N/A | N |
| VOT [21] | Arbitrary | User-specified in the first frame | N | N | 4 | Y |
| RVOT [32] | Arbitrary | User-specified in the first frame | N | N | 5 | Y |
| MPOT | Arbitrary | User-specified in the first frame | Y | Y | 8 | Y |
| RGB+Depth | RGBDT [91] | Arbitrary | User-specified in the first frame | Y | N | 4 | Y |
| 6DOP [50] | Limited by training set | Detected | Y | Y | 6 | Y |
| 6DPT [81] | Arbitrary | User-specified in the first frame | N | Y | 6 | Y |
| RGB+LIDAR | 3dOD [35] | Limited by training set | Detected | Y | N | 7 | Y |
| 3dMOT [84] | Limited, usually vehicles | Detected | Y | N | 7 | Y |
",
+ "bbox": [
+ 76,
+ 108,
+ 890,
+ 263
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "segmentation (VOS), an alternative that introduces masks at the pixel level, MPOT is a more challenging task. Because MPOT provides the matched correspondence for each pixel within the object region across frames [60] (e.g., colored points in Fig. 1(a)&(c)), which makes it possible for applications that require positional information like texture mapping [24]. Nevertheless, VOS that tracks the target area instead of per-pixel location can hardly achieve it. 2) Occlusion is another challenge that comes with MPOT. Not only the one in POT that manually occludes the camera, MPOT but introduces the occlusion raised by the layered position of multiple targets relative to the camera [77, 96] (see Fig. 1(c)). Besides, the occlusion is more complex than the ones in multiple object tracking (MOT). As discussed above, when occlusion occurs, MPOT estimates the pixel correspondence controlled by homography matrix $\\mathbf{H}$ [3], which tends to be sensitive and have a high condition number that can reach up to $5\\mathrm{e}^7$ [95]. This means that a tiny movement of the invisible part is extremely difficult to track.",
+ "bbox": [
+ 75,
+ 277,
+ 472,
+ 580
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "To address the aforementioned limitations, our work draws the inspiration that humans consolidate the visible together with invisible parts of the target for tracking [96]. We propose a tracking framework comprised of procedures for appearance Perception and occlusion Reasoning, namely PRTrack. To track high DoF planar objects, we reformulate the problem of estimating homography matrix $\\mathbf{H}$ as predicting the mask and ordered vertexes for each planar object. With the high-dimensional mask, we can accurately locate the target area at the pixel level. Besides, the ordered vertexes provide per-pixel correspondence across frames for tracking the pose of planar objects. Therefore, in the stage of appearance perception, we propose a dual branch network to predict the ordered vertexes and masks of planar objects based on the historical visible information. For vertexes, we design an encoder with shifted sampling strategy based on the constraint that the vertexes always have a clockwise order. For mask, we aggregate the probability of multiple planar objects with a multi-layered layout, that is, a stack of occluders and occludees. Further, to solve the cases of complex occlusion, we develop an occlusion area",
+ "bbox": [
+ 75,
+ 583,
+ 472,
+ 902
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "localization strategy to indicate the occluded part, by storing the movement of each planar object (i.e., historical H). To be specific, we factorize the sensitive homography matrix, which describes the relative movement of planar objects, into parameters of transition, rotation, and pose. Finally, with the prediction from the mentioned stages, we propose a two-stream self-attention network to jointly refine the predicted planar objects.",
+ "bbox": [
+ 496,
+ 277,
+ 893,
+ 398
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Besides, since there is no available dataset in this field, we build a large-scale benchmark dataset, namely MPOT-3K. Specifically, we shoot 356 videos with 3,717 planar objects and 687,417 annotations. The videos are collected under 9 motion patterns, where the relative movement and occlusion are also included to simulate the real-world scene.",
+ "bbox": [
+ 496,
+ 398,
+ 893,
+ 489
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "The contributions of this paper are two-fold: 1) We collect and annotate the first large-scale benchmark dataset for MPOT, where planar objects are diversely collected expelling bias. Our dataset will be released and boost research in this field. 2) We propose a tracking framework with unified motion and appearance models, which can accurately predict the pose and location of planar objects. Extensive experiments demonstrate the superiority of PRTrack against state-of-the-art approaches.",
+ "bbox": [
+ 496,
+ 489,
+ 893,
+ 627
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2. Related Work",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 640,
+ 640,
+ 656
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "2.1. Planar Object Tracking",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 665,
+ 718,
+ 681
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Benchmark Datasets. Current works focus on single planar object tracking and propose several POT datasets [19, 42, 43, 45, 70, 86]. Metaio [45] is the first dataset collected in the lab and the videos are collected by a monitored camera. Similarly, TMT [70] also shoots videos from the lab, where the annotations are automatically generated by aggregating the results of three planar object trackers. Due to the laboratory setting of the collection, videos in these datasets have similar backgrounds. Thus, POT210 [43] collects video data in the wild under seven motion patterns and encourages the diversity of the recorded videos with complex backgrounds. However, existing benchmarks are insufficient to mimic real-world settings where multiple targets exist. Therefore, our work extends the task into MPOT that",
+ "bbox": [
+ 496,
+ 688,
+ 893,
+ 900
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "23461",
+ "bbox": [
+ 478,
+ 944,
+ 517,
+ 957
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "table",
+ "img_path": "images/4e648d69ebb31962933fe929808ded151752e776c9eb693cfdefcebb991b869c.jpg",
+ "table_caption": [
+ "Table 2: Statistics of tracking datasets. \"I.\" and \"O.\" indicate indoor and outdoor, respectively. \"#\" denotes \"the number of\"."
+ ],
+ "table_footnote": [],
+ "table_body": "| Methods | Traditional | Deep | Ours |
| Motion Patterns | Metrics | CMT [61] | NCC [71] | CCRE [76] | MI [17] | GOP [8] | Gracker [79] | STM [62] | PoST [60] | LISRD [64] | SMask [78] | SRPN+ [38] | SPoint [16] | SOS [73] | GIFT [53] | HDN [95] |
| Overall | \\( S_{0.8} \\uparrow \\) | 09.50 | 00.19 | 09.99 | 26.70 | 56.03 | 67.92 | 09.07 | 33.52 | 63.43 | 64.04 | 67.23 | 71.82 | 72.91 | 75.86 | 76.40 | 82.51 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 64.27 | 61.95 | 34.34 | 37.82 | 62.28 | 88.38 | 52.05 | 39.55 | 72.39 | 73.29 | 75.37 | 71.17 | 70.06 | 72.99 | 79.50 | 94.59 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 37.05 | 03.64 | 14.12 | 23.57 | 45.25 | 85.67 | 25.64 | 25.02 | 56.23 | 58.38 | 60.70 | 54.58 | 53.23 | 56.41 | 65.21 | 92.85 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 10.24 | 00.55 | 19.08 | 38.60 | 62.72 | 78.25 | 23.07 | 40.74 | 80.40 | 69.55 | 74.45 | 81.25 | 85.37 | 84.02 | 82.56 | 90.78 |
| \\( D_{\\epsilon} \\downarrow \\) | 12.30 | 42.15 | 17.27 | 08.68 | 06.37 | 05.23 | 18.79 | 22.71 | 06.93 | 17.06 | 16.77 | 07.67 | 07.76 | 07.80 | 07.44 | 05.07 |
| Moving Occlusion | \\( S_{0.8} \\uparrow \\) | 11.34 | 00.00 | 10.69 | 26.47 | 55.85 | 63.68 | 08.78 | 18.17 | 73.79 | 63.29 | 70.62 | 70.91 | 73.42 | 72.92 | 74.96 | 80.29 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 67.25 | 62.78 | 56.21 | 68.10 | 81.41 | 92.93 | 55.49 | 48.04 | 91.74 | 81.46 | 83.27 | 89.92 | 92.28 | 89.74 | 88.74 | 93.16 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 53.48 | 4.92 | 33.56 | 51.82 | 70.05 | 76.30 | 29.63 | 25.45 | 86.37 | 74.57 | 75.94 | 84.73 | 84.35 | 84.21 | 82.00 | 90.95 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 13.47 | 00.64 | 32.01 | 61.57 | 77.30 | 81.98 | 24.37 | 28.98 | 80.70 | 67.62 | 72.97 | 85.39 | 85.03 | 85.50 | 84.86 | 88.27 |
| \\( D_{\\epsilon} \\downarrow \\) | 06.31 | 42.14 | 21.29 | 08.79 | 04.13 | 07.92 | 16.54 | 27.26 | 06.41 | 15.57 | 14.29 | 06.06 | 06.17 | 06.05 | 06.24 | 05.34 |
| Moving Objects | \\( S_{0.8} \\uparrow \\) | 09.34 | 00.69 | 11.23 | 24.29 | 52.21 | 61.58 | 08.09 | 35.52 | 67.79 | 62.95 | 65.49 | 65.50 | 70.36 | 67.02 | 70.09 | 79.12 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 69.22 | 63.86 | 40.78 | 47.41 | 68.96 | 92.54 | 53.77 | 41.96 | 76.88 | 80.33 | 82.72 | 78.44 | 69.29 | 78.59 | 81.05 | 94.29 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 60.28 | 17.10 | 19.24 | 32.57 | 52.50 | 90.48 | 28.33 | 27.13 | 60.87 | 68.62 | 72.09 | 67.75 | 52.41 | 67.97 | 67.18 | 92.27 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 22.87 | 2.19 | 24.27 | 53.95 | 72.51 | 56.77 | 25.28 | 43.96 | 87.73 | 75.75 | 79.22 | 67.66 | 86.09 | 67.79 | 84.43 | 90.44 |
| \\( D_{\\epsilon} \\downarrow \\) | 12.71 | 43.21 | 17.66 | 06.85 | 05.71 | 04.11 | 21.77 | 19.93 | 05.15 | 14.92 | 15.48 | 04.21 | 04.53 | 04.16 | 05.63 | 05.46 |
| Camera Occlusion | \\( S_{0.8} \\uparrow \\) | 14.61 | 00.00 | 13.31 | 41.79 | 63.59 | 66.41 | 13.80 | 40.21 | 85.46 | 62.46 | 59.40 | 88.85 | 88.28 | 89.43 | 84.01 | 92.46 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 63.54 | 61.93 | 34.28 | 40.28 | 61.72 | 87.98 | 51.72 | 37.15 | 74.06 | 61.66 | 68.71 | 72.18 | 71.90 | 72.10 | 84.51 | 95.85 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 36.89 | 00.38 | 18.42 | 27.84 | 45.04 | 88.24 | 24.59 | 25.31 | 56.76 | 44.83 | 52.31 | 54.67 | 54.42 | 54.60 | 73.16 | 95.84 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 13.19 | 00.06 | 28.32 | 49.74 | 67.38 | 73.75 | 21.70 | 45.37 | 84.32 | 64.86 | 70.75 | 87.57 | 87.15 | 87.50 | 84.52 | 91.53 |
| \\( D_{\\epsilon} \\downarrow \\) | 10.31 | 29.38 | 20.10 | 07.71 | 06.73 | 04.63 | 25.05 | 20.08 | 04.82 | 14.46 | 14.45 | 04.81 | 04.62 | 04.73 | 05.18 | 04.59 |
| Speed↑ | 00.58 | 01.98 | 00.31 | 01.25 | 01.00 | 03.14 | 35.44 | 11.90 | 08.63 | 31.52 | 10.87 | 09.32 | 12.51 | 26.62 | 16.37 | 12.64 |
",
+ "bbox": [
+ 80,
+ 109,
+ 893,
+ 417
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "ation. To fairly compare along with the trajectory level, we adopt the success rate $(S_{\\alpha})$ with the distance threshold $\\alpha$ as",
+ "bbox": [
+ 76,
+ 436,
+ 470,
+ 465
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nS _ {\\alpha} = \\frac {| \\mathbb {C} |}{| \\mathbb {I} |}, \\mathbb {C} = \\left\\{i _ {n} \\in \\mathbb {I} \\right| \\frac {\\sum_ {t = 1} ^ {T} \\left| T _ {\\epsilon} ^ {t} \\right|}{\\sum_ {t = 1} ^ {T} \\left| G T _ {t} \\right|} > \\alpha \\}, \\tag {11}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 112,
+ 477,
+ 468,
+ 516
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "where $\\mathbb{I}$ is the set of planar objects in the video.",
+ "bbox": [
+ 76,
+ 520,
+ 390,
+ 535
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5.2. Implementation Details",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 545,
+ 294,
+ 560
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We compare our model with fifteen representative methods. These methods are implemented by official code or open-source libraries [6, 72]. We run evaluated single object trackers multiple times for different planar objects. And we equip them with the data association strategy based on [12], for integrating each single object tracker into tracking multiple targets online. Besides, we additionally give the reference frames of ground truth if the single object trackers fail to track. Different from other box-based trackers, mask-based trackers (e.g., STM) directly outputs the mask. Thus, we use a rotated box estimator [88] to transfer the mask. We implement PRTrack using two RTX3090 GPUs. In the training stage, we use data augmentations, including flip, pepper noise, rotation, contrast jittering, and perspective transformation. We adopt a data sampling strategy following [41]. We adopt Smooth L1 loss and Cross-Entropy loss for optimizing the predicted vertex heatmaps and masks. Similar to [62], we pretrain SPB on YouTubeVOS [89] and DAVIS [66] datasets. During inference, we use the ordered vertexes to build the planar objects and estimate the homography matrix. We set the threshold as 0.9 in the reuse gate by grid search and set $r$ in Equ. 2 as 0 during inference.",
+ "bbox": [
+ 76,
+ 569,
+ 468,
+ 900
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5.3. Results on MPOT-3K",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 435,
+ 700,
+ 450
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We compare our PRTrack with the other fifteen trackers on MPOT-3K, which contains six traditional POT trackers, four deep-based generic object trackers, and five deep-based planar object trackers. Tab. 3 reports the overall performance as well as the one in the challenging situations of moving occlusion, moving objects, and camera occlusion.",
+ "bbox": [
+ 496,
+ 460,
+ 890,
+ 550
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Deep vs. Traditional Trackers. We can observe that the deep-based trackers have increased by $82.6\\%$ on average, on five metrics for overall performance. This is because deep-based trackers avoid taking some assumptions under the laboratory environment, such as the invariance to illumination and appearance change. Note that Gracker performs best in traditional trackers and achieves competitive performance with the deep-based tracker group. However, since MPOT-3K has many small planar objects, it can not extract sufficient information, thus resulting in suboptimal performance (i.e., $10.3\\%$ degradation in comparison with ours). In terms of deep-based trackers, STM predicts segmentation masks at the pixel level while ignoring the order of vertexes. Thus, it results in inferior performance, where only $9.0\\%$ of targets are correctly tracked.",
+ "bbox": [
+ 496,
+ 551,
+ 890,
+ 777
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Planar Object Trackers vs. Generic Object Trackers. Planar object trackers achieve higher performance (58.9% improvement) against generic object trackers, including SRPN+, SMask, STM, and PoST. A predominant reason is that the generic trackers are subject to the assumption of affine transformation, where only the coarse box estimation is required. Therefore, a powerful planar object tracker should be developed in this field.",
+ "bbox": [
+ 496,
+ 779,
+ 890,
+ 900
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "23466",
+ "bbox": [
+ 478,
+ 945,
+ 519,
+ 955
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/9f635b7d3f32b0622c8b4a686a297d059450fb855b97c6f5a5256412428d9ae6.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 78,
+ 56,
+ 271,
+ 183
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/5a567515146fced29c4d3cd7972f5a77d3892cde57c2ae027f18846aa3631cd2.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 272,
+ 56,
+ 468,
+ 184
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/2826847f90906e62103655908e4f5cf24e3dca71bebb4f4b1a88913615ef853e.jpg",
+ "image_caption": [
+ "(a) POT210",
+ "(b) POT280",
+ "(c) Challenging factors",
+ "Figure 6: Comparison on POT210 and POT280. The averaged Precision on two datasets with different thresholds are reported in (a) and (b), respectively. (c) further shows the performance under seven challenging factors. The data is extracted from [95]."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 145,
+ 185,
+ 393,
+ 306
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "PRTrack vs. SOTA Trackers. PRTrack outperforms other methods with an average improvement of $6.4\\%$ . We also observe a large gain when there exists relative movement among multiple targets. To be specific, when facing challenging situations like moving objects, moving occlusion, and camera occlusion, PRTrack achieves a performance gain of $16.9\\%$ against other advanced algorithms.",
+ "bbox": [
+ 75,
+ 393,
+ 468,
+ 500
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "5.4. Results on POT210 and POT280",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 508,
+ 362,
+ 523
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "To verify the effectiveness of PRTrack on the traditional POT task, we conduct experiments on the POT210 and POT280 datasets, and compare PRTrack with eleven POT methods as shown in Fig. 6. Note that we do not fine-tune PRTrack on these two datasets, again verifying the generalization ability of PRTrack towards unseen scenes. Following [95], the averaged Precision is employed to evaluate all the methods. As the distance threshold $E_{AL}$ increases, the precision tends to be stable when $E_{AL} > 10$ . We notice that PRTrack achieves competitive performance in comparison with advanced algorithms. Besides, PRTrack shows its superiority when handling challenging cases such as occlusion and motion blur.",
+ "bbox": [
+ 75,
+ 532,
+ 468,
+ 729
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "5.5. Ablation Study",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 739,
+ 230,
+ 757
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We conduct ablation studies on MPOT-3K in Tab. 4 to explore the effectiveness of each component in PRTrack. Our baseline is the appearance perception module, without shifted encoder and auxiliary mask. For simplicity, we use \"i&j\" to denote the comparison between the $i^{\\text{th}}$ line and the $j^{\\text{th}}$ line in the rest of this subsection.",
+ "bbox": [
+ 75,
+ 763,
+ 468,
+ 853
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We first perform module-wise ablation studies in the highlighted lines, we observe that PRTrack obtains the largest performance gain of $11.6\\%$ (i.e., $3 \\& 9$ and $4 \\& 10$ )",
+ "bbox": [
+ 75,
+ 854,
+ 468,
+ 901
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Table 4: Ablation study of components and their variants of PRTrack on the validation split of MPOT-3K. The lines with background indicate the module-level ablation study. Here AP, OR, and RT are short of appearance perception, occlusion reasoning, and memory pool. SE and Fac denote the shifted encoder and factorized homography matrix. Msk represents the plane structure branch with mask information. Here, $*$ , $\\bullet$ , and $\\diamond$ are the fusion strategy (Fus), representing concatenation, attention-based, and indicator-based. Gat indicates the reuse gate.",
+ "bbox": [
+ 498,
+ 58,
+ 890,
+ 176
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "table",
+ "img_path": "images/336ceffd2d68e10f12d78a1006a85fc357582769ed336b5be31f14d1b1a96b09.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Modality | Task | Object | Scenario |
| Object category | Object selection | Multiple objects | Pixel correspondence | DoF | Occlusion |
| RGB | MOT [83] | Limited, usually pedestrians or vehicles | Detected | Y | N | 4 | Y |
| MOTS [75] | Limited, usually pedestrians or vehicles | Detected | Y | N | N/A | N |
| VOS [66] | Arbitrary | User-specified in the first frame | N | N | N/A | N |
| VOT [21] | Arbitrary | User-specified in the first frame | N | N | 4 | Y |
| RVOT [32] | Arbitrary | User-specified in the first frame | N | N | 5 | Y |
| MPOT | Arbitrary | User-specified in the first frame | Y | Y | 8 | Y |
| RGB+Depth | RGBDT [91] | Arbitrary | User-specified in the first frame | Y | N | 4 | Y |
| 6DOP [50] | Limited by training set | Detected | Y | Y | 6 | Y |
| 6DPT [81] | Arbitrary | User-specified in the first frame | N | Y | 6 | Y |
| RGB+LIDAR | 3dOD [35] | Limited by training set | Detected | Y | N | 7 | Y |
| 3dMOT [84] | Limited, usually vehicles | Detected | Y | N | 7 | Y |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.279,
+ 0.473,
+ 0.581
+ ],
+ "angle": 0,
+ "content": "segmentation (VOS), an alternative that introduces masks at the pixel level, MPOT is a more challenging task. Because MPOT provides the matched correspondence for each pixel within the object region across frames [60] (e.g., colored points in Fig. 1(a)&(c)), which makes it possible for applications that require positional information like texture mapping [24]. Nevertheless, VOS that tracks the target area instead of per-pixel location can hardly achieve it. 2) Occlusion is another challenge that comes with MPOT. Not only the one in POT that manually occludes the camera, MPOT but introduces the occlusion raised by the layered position of multiple targets relative to the camera [77, 96] (see Fig. 1(c)). Besides, the occlusion is more complex than the ones in multiple object tracking (MOT). As discussed above, when occlusion occurs, MPOT estimates the pixel correspondence controlled by homography matrix \\(\\mathbf{H}\\) [3], which tends to be sensitive and have a high condition number that can reach up to \\(5\\mathrm{e}^7\\) [95]. This means that a tiny movement of the invisible part is extremely difficult to track."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.584,
+ 0.473,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "To address the aforementioned limitations, our work draws the inspiration that humans consolidate the visible together with invisible parts of the target for tracking [96]. We propose a tracking framework comprised of procedures for appearance Perception and occlusion Reasoning, namely PRTrack. To track high DoF planar objects, we reformulate the problem of estimating homography matrix \\(\\mathbf{H}\\) as predicting the mask and ordered vertexes for each planar object. With the high-dimensional mask, we can accurately locate the target area at the pixel level. Besides, the ordered vertexes provide per-pixel correspondence across frames for tracking the pose of planar objects. Therefore, in the stage of appearance perception, we propose a dual branch network to predict the ordered vertexes and masks of planar objects based on the historical visible information. For vertexes, we design an encoder with shifted sampling strategy based on the constraint that the vertexes always have a clockwise order. For mask, we aggregate the probability of multiple planar objects with a multi-layered layout, that is, a stack of occluders and occludees. Further, to solve the cases of complex occlusion, we develop an occlusion area"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.279,
+ 0.895,
+ 0.4
+ ],
+ "angle": 0,
+ "content": "localization strategy to indicate the occluded part, by storing the movement of each planar object (i.e., historical H). To be specific, we factorize the sensitive homography matrix, which describes the relative movement of planar objects, into parameters of transition, rotation, and pose. Finally, with the prediction from the mentioned stages, we propose a two-stream self-attention network to jointly refine the predicted planar objects."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.4,
+ 0.895,
+ 0.49
+ ],
+ "angle": 0,
+ "content": "Besides, since there is no available dataset in this field, we build a large-scale benchmark dataset, namely MPOT-3K. Specifically, we shoot 356 videos with 3,717 planar objects and 687,417 annotations. The videos are collected under 9 motion patterns, where the relative movement and occlusion are also included to simulate the real-world scene."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.49,
+ 0.895,
+ 0.628
+ ],
+ "angle": 0,
+ "content": "The contributions of this paper are two-fold: 1) We collect and annotate the first large-scale benchmark dataset for MPOT, where planar objects are diversely collected expelling bias. Our dataset will be released and boost research in this field. 2) We propose a tracking framework with unified motion and appearance models, which can accurately predict the pose and location of planar objects. Extensive experiments demonstrate the superiority of PRTrack against state-of-the-art approaches."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.641,
+ 0.642,
+ 0.657
+ ],
+ "angle": 0,
+ "content": "2. Related Work"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.666,
+ 0.72,
+ 0.683
+ ],
+ "angle": 0,
+ "content": "2.1. Planar Object Tracking"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.689,
+ 0.895,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Benchmark Datasets. Current works focus on single planar object tracking and propose several POT datasets [19, 42, 43, 45, 70, 86]. Metaio [45] is the first dataset collected in the lab and the videos are collected by a monitored camera. Similarly, TMT [70] also shoots videos from the lab, where the annotations are automatically generated by aggregating the results of three planar object trackers. Due to the laboratory setting of the collection, videos in these datasets have similar backgrounds. Thus, POT210 [43] collects video data in the wild under seven motion patterns and encourages the diversity of the recorded videos with complex backgrounds. However, existing benchmarks are insufficient to mimic real-world settings where multiple targets exist. Therefore, our work extends the task into MPOT that"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.479,
+ 0.945,
+ 0.518,
+ 0.958
+ ],
+ "angle": 0,
+ "content": "23461"
+ }
+ ],
+ [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.078,
+ 0.09,
+ 0.47,
+ 0.117
+ ],
+ "angle": 0,
+ "content": "Table 2: Statistics of tracking datasets. \"I.\" and \"O.\" indicate indoor and outdoor, respectively. \"#\" denotes \"the number of\"."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.081,
+ 0.117,
+ 0.468,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "| Methods | Traditional | Deep | Ours |
| Motion Patterns | Metrics | CMT [61] | NCC [71] | CCRE [76] | MI [17] | GOP [8] | Gracker [79] | STM [62] | PoST [60] | LISRD [64] | SMask [78] | SRPN+ [38] | SPoint [16] | SOS [73] | GIFT [53] | HDN [95] |
| Overall | \\( S_{0.8} \\uparrow \\) | 09.50 | 00.19 | 09.99 | 26.70 | 56.03 | 67.92 | 09.07 | 33.52 | 63.43 | 64.04 | 67.23 | 71.82 | 72.91 | 75.86 | 76.40 | 82.51 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 64.27 | 61.95 | 34.34 | 37.82 | 62.28 | 88.38 | 52.05 | 39.55 | 72.39 | 73.29 | 75.37 | 71.17 | 70.06 | 72.99 | 79.50 | 94.59 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 37.05 | 03.64 | 14.12 | 23.57 | 45.25 | 85.67 | 25.64 | 25.02 | 56.23 | 58.38 | 60.70 | 54.58 | 53.23 | 56.41 | 65.21 | 92.85 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 10.24 | 00.55 | 19.08 | 38.60 | 62.72 | 78.25 | 23.07 | 40.74 | 80.40 | 69.55 | 74.45 | 81.25 | 85.37 | 84.02 | 82.56 | 90.78 |
| \\( D_{\\epsilon} \\downarrow \\) | 12.30 | 42.15 | 17.27 | 08.68 | 06.37 | 05.23 | 18.79 | 22.71 | 06.93 | 17.06 | 16.77 | 07.67 | 07.76 | 07.80 | 07.44 | 05.07 |
| Moving Occlusion | \\( S_{0.8} \\uparrow \\) | 11.34 | 00.00 | 10.69 | 26.47 | 55.85 | 63.68 | 08.78 | 18.17 | 73.79 | 63.29 | 70.62 | 70.91 | 73.42 | 72.92 | 74.96 | 80.29 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 67.25 | 62.78 | 56.21 | 68.10 | 81.41 | 92.93 | 55.49 | 48.04 | 91.74 | 81.46 | 83.27 | 89.92 | 92.28 | 89.74 | 88.74 | 93.16 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 53.48 | 4.92 | 33.56 | 51.82 | 70.05 | 76.30 | 29.63 | 25.45 | 86.37 | 74.57 | 75.94 | 84.73 | 84.35 | 84.21 | 82.00 | 90.95 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 13.47 | 00.64 | 32.01 | 61.57 | 77.30 | 81.98 | 24.37 | 28.98 | 80.70 | 67.62 | 72.97 | 85.39 | 85.03 | 85.50 | 84.86 | 88.27 |
| \\( D_{\\epsilon} \\downarrow \\) | 06.31 | 42.14 | 21.29 | 08.79 | 04.13 | 07.92 | 16.54 | 27.26 | 06.41 | 15.57 | 14.29 | 06.06 | 06.17 | 06.05 | 06.24 | 05.34 |
| Moving Objects | \\( S_{0.8} \\uparrow \\) | 09.34 | 00.69 | 11.23 | 24.29 | 52.21 | 61.58 | 08.09 | 35.52 | 67.79 | 62.95 | 65.49 | 65.50 | 70.36 | 67.02 | 70.09 | 79.12 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 69.22 | 63.86 | 40.78 | 47.41 | 68.96 | 92.54 | 53.77 | 41.96 | 76.88 | 80.33 | 82.72 | 78.44 | 69.29 | 78.59 | 81.05 | 94.29 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 60.28 | 17.10 | 19.24 | 32.57 | 52.50 | 90.48 | 28.33 | 27.13 | 60.87 | 68.62 | 72.09 | 67.75 | 52.41 | 67.97 | 67.18 | 92.27 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 22.87 | 2.19 | 24.27 | 53.95 | 72.51 | 56.77 | 25.28 | 43.96 | 87.73 | 75.75 | 79.22 | 67.66 | 86.09 | 67.79 | 84.43 | 90.44 |
| \\( D_{\\epsilon} \\downarrow \\) | 12.71 | 43.21 | 17.66 | 06.85 | 05.71 | 04.11 | 21.77 | 19.93 | 05.15 | 14.92 | 15.48 | 04.21 | 04.53 | 04.16 | 05.63 | 05.46 |
| Camera Occlusion | \\( S_{0.8} \\uparrow \\) | 14.61 | 00.00 | 13.31 | 41.79 | 63.59 | 66.41 | 13.80 | 40.21 | 85.46 | 62.46 | 59.40 | 88.85 | 88.28 | 89.43 | 84.01 | 92.46 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 63.54 | 61.93 | 34.28 | 40.28 | 61.72 | 87.98 | 51.72 | 37.15 | 74.06 | 61.66 | 68.71 | 72.18 | 71.90 | 72.10 | 84.51 | 95.85 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 36.89 | 00.38 | 18.42 | 27.84 | 45.04 | 88.24 | 24.59 | 25.31 | 56.76 | 44.83 | 52.31 | 54.67 | 54.42 | 54.60 | 73.16 | 95.84 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 13.19 | 00.06 | 28.32 | 49.74 | 67.38 | 73.75 | 21.70 | 45.37 | 84.32 | 64.86 | 70.75 | 87.57 | 87.15 | 87.50 | 84.52 | 91.53 |
| \\( D_{\\epsilon} \\downarrow \\) | 10.31 | 29.38 | 20.10 | 07.71 | 06.73 | 04.63 | 25.05 | 20.08 | 04.82 | 14.46 | 14.45 | 04.81 | 04.62 | 04.73 | 05.18 | 04.59 |
| Speed↑ | 00.58 | 01.98 | 00.31 | 01.25 | 01.00 | 03.14 | 35.44 | 11.90 | 08.63 | 31.52 | 10.87 | 09.32 | 12.51 | 26.62 | 16.37 | 12.64 |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.437,
+ 0.471,
+ 0.467
+ ],
+ "angle": 0,
+ "content": "ation. To fairly compare along with the trajectory level, we adopt the success rate \\((S_{\\alpha})\\) with the distance threshold \\(\\alpha\\) as"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.113,
+ 0.478,
+ 0.469,
+ 0.517
+ ],
+ "angle": 0,
+ "content": "\\[\nS _ {\\alpha} = \\frac {| \\mathbb {C} |}{| \\mathbb {I} |}, \\mathbb {C} = \\left\\{i _ {n} \\in \\mathbb {I} \\right| \\frac {\\sum_ {t = 1} ^ {T} \\left| T _ {\\epsilon} ^ {t} \\right|}{\\sum_ {t = 1} ^ {T} \\left| G T _ {t} \\right|} > \\alpha \\}, \\tag {11}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.078,
+ 0.521,
+ 0.391,
+ 0.536
+ ],
+ "angle": 0,
+ "content": "where \\(\\mathbb{I}\\) is the set of planar objects in the video."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.546,
+ 0.295,
+ 0.561
+ ],
+ "angle": 0,
+ "content": "5.2. Implementation Details"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.57,
+ 0.47,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "We compare our model with fifteen representative methods. These methods are implemented by official code or open-source libraries [6, 72]. We run evaluated single object trackers multiple times for different planar objects. And we equip them with the data association strategy based on [12], for integrating each single object tracker into tracking multiple targets online. Besides, we additionally give the reference frames of ground truth if the single object trackers fail to track. Different from other box-based trackers, mask-based trackers (e.g., STM) directly outputs the mask. Thus, we use a rotated box estimator [88] to transfer the mask. We implement PRTrack using two RTX3090 GPUs. In the training stage, we use data augmentations, including flip, pepper noise, rotation, contrast jittering, and perspective transformation. We adopt a data sampling strategy following [41]. We adopt Smooth L1 loss and Cross-Entropy loss for optimizing the predicted vertex heatmaps and masks. Similar to [62], we pretrain SPB on YouTubeVOS [89] and DAVIS [66] datasets. During inference, we use the ordered vertexes to build the planar objects and estimate the homography matrix. We set the threshold as 0.9 in the reuse gate by grid search and set \\( r \\) in Equ. 2 as 0 during inference."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.436,
+ 0.702,
+ 0.451
+ ],
+ "angle": 0,
+ "content": "5.3. Results on MPOT-3K"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.461,
+ 0.892,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "We compare our PRTrack with the other fifteen trackers on MPOT-3K, which contains six traditional POT trackers, four deep-based generic object trackers, and five deep-based planar object trackers. Tab. 3 reports the overall performance as well as the one in the challenging situations of moving occlusion, moving objects, and camera occlusion."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.553,
+ 0.892,
+ 0.779
+ ],
+ "angle": 0,
+ "content": "Deep vs. Traditional Trackers. We can observe that the deep-based trackers have increased by \\(82.6\\%\\) on average, on five metrics for overall performance. This is because deep-based trackers avoid taking some assumptions under the laboratory environment, such as the invariance to illumination and appearance change. Note that Gracker performs best in traditional trackers and achieves competitive performance with the deep-based tracker group. However, since MPOT-3K has many small planar objects, it can not extract sufficient information, thus resulting in suboptimal performance (i.e., \\(10.3\\%\\) degradation in comparison with ours). In terms of deep-based trackers, STM predicts segmentation masks at the pixel level while ignoring the order of vertexes. Thus, it results in inferior performance, where only \\(9.0\\%\\) of targets are correctly tracked."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.78,
+ 0.892,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Planar Object Trackers vs. Generic Object Trackers. Planar object trackers achieve higher performance (58.9% improvement) against generic object trackers, including SRPN+, SMask, STM, and PoST. A predominant reason is that the generic trackers are subject to the assumption of affine transformation, where only the coarse box estimation is required. Therefore, a powerful planar object tracker should be developed in this field."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.48,
+ 0.946,
+ 0.52,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "23466"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.08,
+ 0.058,
+ 0.272,
+ 0.184
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.14,
+ 0.185,
+ 0.212,
+ 0.197
+ ],
+ "angle": 0,
+ "content": "(a) POT210"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.274,
+ 0.058,
+ 0.47,
+ 0.185
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.335,
+ 0.185,
+ 0.409,
+ 0.197
+ ],
+ "angle": 0,
+ "content": "(b) POT280"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.147,
+ 0.186,
+ 0.395,
+ 0.308
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.221,
+ 0.31,
+ 0.36,
+ 0.323
+ ],
+ "angle": 0,
+ "content": "(c) Challenging factors"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.076,
+ 0.324,
+ 0.471,
+ 0.378
+ ],
+ "angle": 0,
+ "content": "Figure 6: Comparison on POT210 and POT280. The averaged Precision on two datasets with different thresholds are reported in (a) and (b), respectively. (c) further shows the performance under seven challenging factors. The data is extracted from [95]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.394,
+ 0.47,
+ 0.501
+ ],
+ "angle": 0,
+ "content": "PRTrack vs. SOTA Trackers. PRTrack outperforms other methods with an average improvement of \\(6.4\\%\\). We also observe a large gain when there exists relative movement among multiple targets. To be specific, when facing challenging situations like moving objects, moving occlusion, and camera occlusion, PRTrack achieves a performance gain of \\(16.9\\%\\) against other advanced algorithms."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.51,
+ 0.364,
+ 0.525
+ ],
+ "angle": 0,
+ "content": "5.4. Results on POT210 and POT280"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.534,
+ 0.47,
+ 0.73
+ ],
+ "angle": 0,
+ "content": "To verify the effectiveness of PRTrack on the traditional POT task, we conduct experiments on the POT210 and POT280 datasets, and compare PRTrack with eleven POT methods as shown in Fig. 6. Note that we do not fine-tune PRTrack on these two datasets, again verifying the generalization ability of PRTrack towards unseen scenes. Following [95], the averaged Precision is employed to evaluate all the methods. As the distance threshold \\( E_{AL} \\) increases, the precision tends to be stable when \\( E_{AL} > 10 \\). We notice that PRTrack achieves competitive performance in comparison with advanced algorithms. Besides, PRTrack shows its superiority when handling challenging cases such as occlusion and motion blur."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.741,
+ 0.231,
+ 0.758
+ ],
+ "angle": 0,
+ "content": "5.5. Ablation Study"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.764,
+ 0.469,
+ 0.854
+ ],
+ "angle": 0,
+ "content": "We conduct ablation studies on MPOT-3K in Tab. 4 to explore the effectiveness of each component in PRTrack. Our baseline is the appearance perception module, without shifted encoder and auxiliary mask. For simplicity, we use \"i&j\" to denote the comparison between the \\( i^{\\text{th}} \\) line and the \\( j^{\\text{th}} \\) line in the rest of this subsection."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.856,
+ 0.47,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "We first perform module-wise ablation studies in the highlighted lines, we observe that PRTrack obtains the largest performance gain of \\(11.6\\%\\) (i.e., \\(3 \\& 9\\) and \\(4 \\& 10\\))"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.499,
+ 0.059,
+ 0.892,
+ 0.178
+ ],
+ "angle": 0,
+ "content": "Table 4: Ablation study of components and their variants of PRTrack on the validation split of MPOT-3K. The lines with background indicate the module-level ablation study. Here AP, OR, and RT are short of appearance perception, occlusion reasoning, and memory pool. SE and Fac denote the shifted encoder and factorized homography matrix. Msk represents the plane structure branch with mask information. Here, \\(*\\), \\(\\bullet\\), and \\(\\diamond\\) are the fusion strategy (Fus), representing concatenation, attention-based, and indicator-based. Gat indicates the reuse gate."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.504,
+ 0.179,
+ 0.89,
+ 0.322
+ ],
+ "angle": 0,
+ "content": "| Modality | Task | Object | Scenario |
| Object category | Object selection | Multiple objects | Pixel correspondence | DoF | Occlusion |
| RGB | MOT [83] | Limited, usually pedestrians or vehicles | Detected | Y | N | 4 | Y |
| MOTS [75] | Limited, usually pedestrians or vehicles | Detected | Y | N | N/A | N |
| VOS [66] | Arbitrary | User-specified in the first frame | N | N | N/A | N |
| VOT [21] | Arbitrary | User-specified in the first frame | N | N | 4 | Y |
| RVOT [32] | Arbitrary | User-specified in the first frame | N | N | 5 | Y |
| MPOT | Arbitrary | User-specified in the first frame | Y | Y | 8 | Y |
| RGB+Depth | RGBDT [91] | Arbitrary | User-specified in the first frame | Y | N | 4 | Y |
| 6DOP [50] | Limited by training set | Detected | Y | Y | 6 | Y |
| 6DPT [81] | Arbitrary | User-specified in the first frame | N | Y | 6 | Y |
| RGB+LIDAR | 3dOD [35] | Limited by training set | Detected | Y | N | 7 | Y |
| 3dMOT [84] | Limited, usually vehicles | Detected | Y | N | 7 | Y |
+
+segmentation (VOS), an alternative that introduces masks at the pixel level, MPOT is a more challenging task. Because MPOT provides the matched correspondence for each pixel within the object region across frames [60] (e.g., colored points in Fig. 1(a)&(c)), which makes it possible for applications that require positional information like texture mapping [24]. Nevertheless, VOS that tracks the target area instead of per-pixel location can hardly achieve it. 2) Occlusion is another challenge that comes with MPOT. Not only the one in POT that manually occludes the camera, MPOT but introduces the occlusion raised by the layered position of multiple targets relative to the camera [77, 96] (see Fig. 1(c)). Besides, the occlusion is more complex than the ones in multiple object tracking (MOT). As discussed above, when occlusion occurs, MPOT estimates the pixel correspondence controlled by homography matrix $\mathbf{H}$ [3], which tends to be sensitive and have a high condition number that can reach up to $5\mathrm{e}^7$ [95]. This means that a tiny movement of the invisible part is extremely difficult to track.
+
+To address the aforementioned limitations, our work draws the inspiration that humans consolidate the visible together with invisible parts of the target for tracking [96]. We propose a tracking framework comprised of procedures for appearance Perception and occlusion Reasoning, namely PRTrack. To track high DoF planar objects, we reformulate the problem of estimating homography matrix $\mathbf{H}$ as predicting the mask and ordered vertexes for each planar object. With the high-dimensional mask, we can accurately locate the target area at the pixel level. Besides, the ordered vertexes provide per-pixel correspondence across frames for tracking the pose of planar objects. Therefore, in the stage of appearance perception, we propose a dual branch network to predict the ordered vertexes and masks of planar objects based on the historical visible information. For vertexes, we design an encoder with shifted sampling strategy based on the constraint that the vertexes always have a clockwise order. For mask, we aggregate the probability of multiple planar objects with a multi-layered layout, that is, a stack of occluders and occludees. Further, to solve the cases of complex occlusion, we develop an occlusion area
+
+localization strategy to indicate the occluded part, by storing the movement of each planar object (i.e., historical H). To be specific, we factorize the sensitive homography matrix, which describes the relative movement of planar objects, into parameters of transition, rotation, and pose. Finally, with the prediction from the mentioned stages, we propose a two-stream self-attention network to jointly refine the predicted planar objects.
+
+Besides, since there is no available dataset in this field, we build a large-scale benchmark dataset, namely MPOT-3K. Specifically, we shoot 356 videos with 3,717 planar objects and 687,417 annotations. The videos are collected under 9 motion patterns, where the relative movement and occlusion are also included to simulate the real-world scene.
+
+The contributions of this paper are two-fold: 1) We collect and annotate the first large-scale benchmark dataset for MPOT, where planar objects are diversely collected expelling bias. Our dataset will be released and boost research in this field. 2) We propose a tracking framework with unified motion and appearance models, which can accurately predict the pose and location of planar objects. Extensive experiments demonstrate the superiority of PRTrack against state-of-the-art approaches.
+
+# 2. Related Work
+
+# 2.1. Planar Object Tracking
+
+Benchmark Datasets. Current works focus on single planar object tracking and propose several POT datasets [19, 42, 43, 45, 70, 86]. Metaio [45] is the first dataset collected in the lab and the videos are collected by a monitored camera. Similarly, TMT [70] also shoots videos from the lab, where the annotations are automatically generated by aggregating the results of three planar object trackers. Due to the laboratory setting of the collection, videos in these datasets have similar backgrounds. Thus, POT210 [43] collects video data in the wild under seven motion patterns and encourages the diversity of the recorded videos with complex backgrounds. However, existing benchmarks are insufficient to mimic real-world settings where multiple targets exist. Therefore, our work extends the task into MPOT that
+
+Table 2: Statistics of tracking datasets. "I." and "O." indicate indoor and outdoor, respectively. "#" denotes "the number of".
+
+| Methods | Traditional | Deep | Ours |
| Motion Patterns | Metrics | CMT [61] | NCC [71] | CCRE [76] | MI [17] | GOP [8] | Gracker [79] | STM [62] | PoST [60] | LISRD [64] | SMask [78] | SRPN+ [38] | SPoint [16] | SOS [73] | GIFT [53] | HDN [95] |
| Overall | \( S_{0.8} \uparrow \) | 09.50 | 00.19 | 09.99 | 26.70 | 56.03 | 67.92 | 09.07 | 33.52 | 63.43 | 64.04 | 67.23 | 71.82 | 72.91 | 75.86 | 76.40 | 82.51 |
| \( Acc_{\epsilon} \uparrow \) | 64.27 | 61.95 | 34.34 | 37.82 | 62.28 | 88.38 | 52.05 | 39.55 | 72.39 | 73.29 | 75.37 | 71.17 | 70.06 | 72.99 | 79.50 | 94.59 |
| \( Pr_{\epsilon} \uparrow \) | 37.05 | 03.64 | 14.12 | 23.57 | 45.25 | 85.67 | 25.64 | 25.02 | 56.23 | 58.38 | 60.70 | 54.58 | 53.23 | 56.41 | 65.21 | 92.85 |
| \( Rc_{\epsilon} \uparrow \) | 10.24 | 00.55 | 19.08 | 38.60 | 62.72 | 78.25 | 23.07 | 40.74 | 80.40 | 69.55 | 74.45 | 81.25 | 85.37 | 84.02 | 82.56 | 90.78 |
| \( D_{\epsilon} \downarrow \) | 12.30 | 42.15 | 17.27 | 08.68 | 06.37 | 05.23 | 18.79 | 22.71 | 06.93 | 17.06 | 16.77 | 07.67 | 07.76 | 07.80 | 07.44 | 05.07 |
| Moving Occlusion | \( S_{0.8} \uparrow \) | 11.34 | 00.00 | 10.69 | 26.47 | 55.85 | 63.68 | 08.78 | 18.17 | 73.79 | 63.29 | 70.62 | 70.91 | 73.42 | 72.92 | 74.96 | 80.29 |
| \( Acc_{\epsilon} \uparrow \) | 67.25 | 62.78 | 56.21 | 68.10 | 81.41 | 92.93 | 55.49 | 48.04 | 91.74 | 81.46 | 83.27 | 89.92 | 92.28 | 89.74 | 88.74 | 93.16 |
| \( Pr_{\epsilon} \uparrow \) | 53.48 | 4.92 | 33.56 | 51.82 | 70.05 | 76.30 | 29.63 | 25.45 | 86.37 | 74.57 | 75.94 | 84.73 | 84.35 | 84.21 | 82.00 | 90.95 |
| \( Rc_{\epsilon} \uparrow \) | 13.47 | 00.64 | 32.01 | 61.57 | 77.30 | 81.98 | 24.37 | 28.98 | 80.70 | 67.62 | 72.97 | 85.39 | 85.03 | 85.50 | 84.86 | 88.27 |
| \( D_{\epsilon} \downarrow \) | 06.31 | 42.14 | 21.29 | 08.79 | 04.13 | 07.92 | 16.54 | 27.26 | 06.41 | 15.57 | 14.29 | 06.06 | 06.17 | 06.05 | 06.24 | 05.34 |
| Moving Objects | \( S_{0.8} \uparrow \) | 09.34 | 00.69 | 11.23 | 24.29 | 52.21 | 61.58 | 08.09 | 35.52 | 67.79 | 62.95 | 65.49 | 65.50 | 70.36 | 67.02 | 70.09 | 79.12 |
| \( Acc_{\epsilon} \uparrow \) | 69.22 | 63.86 | 40.78 | 47.41 | 68.96 | 92.54 | 53.77 | 41.96 | 76.88 | 80.33 | 82.72 | 78.44 | 69.29 | 78.59 | 81.05 | 94.29 |
| \( Pr_{\epsilon} \uparrow \) | 60.28 | 17.10 | 19.24 | 32.57 | 52.50 | 90.48 | 28.33 | 27.13 | 60.87 | 68.62 | 72.09 | 67.75 | 52.41 | 67.97 | 67.18 | 92.27 |
| \( Rc_{\epsilon} \uparrow \) | 22.87 | 2.19 | 24.27 | 53.95 | 72.51 | 56.77 | 25.28 | 43.96 | 87.73 | 75.75 | 79.22 | 67.66 | 86.09 | 67.79 | 84.43 | 90.44 |
| \( D_{\epsilon} \downarrow \) | 12.71 | 43.21 | 17.66 | 06.85 | 05.71 | 04.11 | 21.77 | 19.93 | 05.15 | 14.92 | 15.48 | 04.21 | 04.53 | 04.16 | 05.63 | 05.46 |
| Camera Occlusion | \( S_{0.8} \uparrow \) | 14.61 | 00.00 | 13.31 | 41.79 | 63.59 | 66.41 | 13.80 | 40.21 | 85.46 | 62.46 | 59.40 | 88.85 | 88.28 | 89.43 | 84.01 | 92.46 |
| \( Acc_{\epsilon} \uparrow \) | 63.54 | 61.93 | 34.28 | 40.28 | 61.72 | 87.98 | 51.72 | 37.15 | 74.06 | 61.66 | 68.71 | 72.18 | 71.90 | 72.10 | 84.51 | 95.85 |
| \( Pr_{\epsilon} \uparrow \) | 36.89 | 00.38 | 18.42 | 27.84 | 45.04 | 88.24 | 24.59 | 25.31 | 56.76 | 44.83 | 52.31 | 54.67 | 54.42 | 54.60 | 73.16 | 95.84 |
| \( Rc_{\epsilon} \uparrow \) | 13.19 | 00.06 | 28.32 | 49.74 | 67.38 | 73.75 | 21.70 | 45.37 | 84.32 | 64.86 | 70.75 | 87.57 | 87.15 | 87.50 | 84.52 | 91.53 |
| \( D_{\epsilon} \downarrow \) | 10.31 | 29.38 | 20.10 | 07.71 | 06.73 | 04.63 | 25.05 | 20.08 | 04.82 | 14.46 | 14.45 | 04.81 | 04.62 | 04.73 | 05.18 | 04.59 |
| Speed↑ | 00.58 | 01.98 | 00.31 | 01.25 | 01.00 | 03.14 | 35.44 | 11.90 | 08.63 | 31.52 | 10.87 | 09.32 | 12.51 | 26.62 | 16.37 | 12.64 |
+
+ation. To fairly compare along with the trajectory level, we adopt the success rate $(S_{\alpha})$ with the distance threshold $\alpha$ as
+
+$$
+S _ {\alpha} = \frac {| \mathbb {C} |}{| \mathbb {I} |}, \mathbb {C} = \left\{i _ {n} \in \mathbb {I} \right| \frac {\sum_ {t = 1} ^ {T} \left| T _ {\epsilon} ^ {t} \right|}{\sum_ {t = 1} ^ {T} \left| G T _ {t} \right|} > \alpha \}, \tag {11}
+$$
+
+where $\mathbb{I}$ is the set of planar objects in the video.
+
+# 5.2. Implementation Details
+
+We compare our model with fifteen representative methods. These methods are implemented by official code or open-source libraries [6, 72]. We run evaluated single object trackers multiple times for different planar objects. And we equip them with the data association strategy based on [12], for integrating each single object tracker into tracking multiple targets online. Besides, we additionally give the reference frames of ground truth if the single object trackers fail to track. Different from other box-based trackers, mask-based trackers (e.g., STM) directly outputs the mask. Thus, we use a rotated box estimator [88] to transfer the mask. We implement PRTrack using two RTX3090 GPUs. In the training stage, we use data augmentations, including flip, pepper noise, rotation, contrast jittering, and perspective transformation. We adopt a data sampling strategy following [41]. We adopt Smooth L1 loss and Cross-Entropy loss for optimizing the predicted vertex heatmaps and masks. Similar to [62], we pretrain SPB on YouTubeVOS [89] and DAVIS [66] datasets. During inference, we use the ordered vertexes to build the planar objects and estimate the homography matrix. We set the threshold as 0.9 in the reuse gate by grid search and set $r$ in Equ. 2 as 0 during inference.
+
+# 5.3. Results on MPOT-3K
+
+We compare our PRTrack with the other fifteen trackers on MPOT-3K, which contains six traditional POT trackers, four deep-based generic object trackers, and five deep-based planar object trackers. Tab. 3 reports the overall performance as well as the one in the challenging situations of moving occlusion, moving objects, and camera occlusion.
+
+Deep vs. Traditional Trackers. We can observe that the deep-based trackers have increased by $82.6\%$ on average, on five metrics for overall performance. This is because deep-based trackers avoid taking some assumptions under the laboratory environment, such as the invariance to illumination and appearance change. Note that Gracker performs best in traditional trackers and achieves competitive performance with the deep-based tracker group. However, since MPOT-3K has many small planar objects, it can not extract sufficient information, thus resulting in suboptimal performance (i.e., $10.3\%$ degradation in comparison with ours). In terms of deep-based trackers, STM predicts segmentation masks at the pixel level while ignoring the order of vertexes. Thus, it results in inferior performance, where only $9.0\%$ of targets are correctly tracked.
+
+Planar Object Trackers vs. Generic Object Trackers. Planar object trackers achieve higher performance (58.9% improvement) against generic object trackers, including SRPN+, SMask, STM, and PoST. A predominant reason is that the generic trackers are subject to the assumption of affine transformation, where only the coarse box estimation is required. Therefore, a powerful planar object tracker should be developed in this field.
+
+
+
+
+
+
+(a) POT210
+(b) POT280
+(c) Challenging factors
+Figure 6: Comparison on POT210 and POT280. The averaged Precision on two datasets with different thresholds are reported in (a) and (b), respectively. (c) further shows the performance under seven challenging factors. The data is extracted from [95].
+
+PRTrack vs. SOTA Trackers. PRTrack outperforms other methods with an average improvement of $6.4\%$ . We also observe a large gain when there exists relative movement among multiple targets. To be specific, when facing challenging situations like moving objects, moving occlusion, and camera occlusion, PRTrack achieves a performance gain of $16.9\%$ against other advanced algorithms.
+
+# 5.4. Results on POT210 and POT280
+
+To verify the effectiveness of PRTrack on the traditional POT task, we conduct experiments on the POT210 and POT280 datasets, and compare PRTrack with eleven POT methods as shown in Fig. 6. Note that we do not fine-tune PRTrack on these two datasets, again verifying the generalization ability of PRTrack towards unseen scenes. Following [95], the averaged Precision is employed to evaluate all the methods. As the distance threshold $E_{AL}$ increases, the precision tends to be stable when $E_{AL} > 10$ . We notice that PRTrack achieves competitive performance in comparison with advanced algorithms. Besides, PRTrack shows its superiority when handling challenging cases such as occlusion and motion blur.
+
+# 5.5. Ablation Study
+
+We conduct ablation studies on MPOT-3K in Tab. 4 to explore the effectiveness of each component in PRTrack. Our baseline is the appearance perception module, without shifted encoder and auxiliary mask. For simplicity, we use "i&j" to denote the comparison between the $i^{\text{th}}$ line and the $j^{\text{th}}$ line in the rest of this subsection.
+
+We first perform module-wise ablation studies in the highlighted lines, we observe that PRTrack obtains the largest performance gain of $11.6\%$ (i.e., $3 \& 9$ and $4 \& 10$ )
+
+Table 4: Ablation study of components and their variants of PRTrack on the validation split of MPOT-3K. The lines with background indicate the module-level ablation study. Here AP, OR, and RT are short of appearance perception, occlusion reasoning, and memory pool. SE and Fac denote the shifted encoder and factorized homography matrix. Msk represents the plane structure branch with mask information. Here, $*$ , $\bullet$ , and $\diamond$ are the fusion strategy (Fus), representing concatenation, attention-based, and indicator-based. Gat indicates the reuse gate.
+
+| Modality | Task | Object | Scenario |
| Object category | Object selection | Multiple objects | Pixel correspondence | DoF | Occlusion |
| RGB | MOT [83] | Limited, usually pedestrians or vehicles | Detected | Y | N | 4 | Y |
| MOTS [75] | Limited, usually pedestrians or vehicles | Detected | Y | N | N/A | N |
| VOS [66] | Arbitrary | User-specified in the first frame | N | N | N/A | N |
| VOT [21] | Arbitrary | User-specified in the first frame | N | N | 4 | Y |
| RVOT [32] | Arbitrary | User-specified in the first frame | N | N | 5 | Y |
| MPOT | Arbitrary | User-specified in the first frame | Y | Y | 8 | Y |
| RGB+Depth | RGBDT [91] | Arbitrary | User-specified in the first frame | Y | N | 4 | Y |
| 6DOP [50] | Limited by training set | Detected | Y | Y | 6 | Y |
| 6DPT [81] | Arbitrary | User-specified in the first frame | N | Y | 6 | Y |
| RGB+LIDAR | 3dOD [35] | Limited by training set | Detected | Y | N | 7 | Y |
| 3dMOT [84] | Limited, usually vehicles | Detected | Y | N | 7 | Y |
",
+ "image_path": "6df07e4fc3e37da7ff6ee7f1d7c5ef3e88a740ed266797426aa31da7119039c2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 46,
+ 220,
+ 289,
+ 460
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 220,
+ 289,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 220,
+ 289,
+ 460
+ ],
+ "type": "text",
+ "content": "segmentation (VOS), an alternative that introduces masks at the pixel level, MPOT is a more challenging task. Because MPOT provides the matched correspondence for each pixel within the object region across frames [60] (e.g., colored points in Fig. 1(a)&(c)), which makes it possible for applications that require positional information like texture mapping [24]. Nevertheless, VOS that tracks the target area instead of per-pixel location can hardly achieve it. 2) Occlusion is another challenge that comes with MPOT. Not only the one in POT that manually occludes the camera, MPOT but introduces the occlusion raised by the layered position of multiple targets relative to the camera [77, 96] (see Fig. 1(c)). Besides, the occlusion is more complex than the ones in multiple object tracking (MOT). As discussed above, when occlusion occurs, MPOT estimates the pixel correspondence controlled by homography matrix "
+ },
+ {
+ "bbox": [
+ 46,
+ 220,
+ 289,
+ 460
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{H}"
+ },
+ {
+ "bbox": [
+ 46,
+ 220,
+ 289,
+ 460
+ ],
+ "type": "text",
+ "content": " [3], which tends to be sensitive and have a high condition number that can reach up to "
+ },
+ {
+ "bbox": [
+ 46,
+ 220,
+ 289,
+ 460
+ ],
+ "type": "inline_equation",
+ "content": "5\\mathrm{e}^7"
+ },
+ {
+ "bbox": [
+ 46,
+ 220,
+ 289,
+ 460
+ ],
+ "type": "text",
+ "content": " [95]. This means that a tiny movement of the invisible part is extremely difficult to track."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 46,
+ 462,
+ 289,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 462,
+ 289,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 462,
+ 289,
+ 715
+ ],
+ "type": "text",
+ "content": "To address the aforementioned limitations, our work draws the inspiration that humans consolidate the visible together with invisible parts of the target for tracking [96]. We propose a tracking framework comprised of procedures for appearance Perception and occlusion Reasoning, namely PRTrack. To track high DoF planar objects, we reformulate the problem of estimating homography matrix "
+ },
+ {
+ "bbox": [
+ 46,
+ 462,
+ 289,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{H}"
+ },
+ {
+ "bbox": [
+ 46,
+ 462,
+ 289,
+ 715
+ ],
+ "type": "text",
+ "content": " as predicting the mask and ordered vertexes for each planar object. With the high-dimensional mask, we can accurately locate the target area at the pixel level. Besides, the ordered vertexes provide per-pixel correspondence across frames for tracking the pose of planar objects. Therefore, in the stage of appearance perception, we propose a dual branch network to predict the ordered vertexes and masks of planar objects based on the historical visible information. For vertexes, we design an encoder with shifted sampling strategy based on the constraint that the vertexes always have a clockwise order. For mask, we aggregate the probability of multiple planar objects with a multi-layered layout, that is, a stack of occluders and occludees. Further, to solve the cases of complex occlusion, we develop an occlusion area"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 304,
+ 220,
+ 547,
+ 316
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 220,
+ 547,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 220,
+ 547,
+ 316
+ ],
+ "type": "text",
+ "content": "localization strategy to indicate the occluded part, by storing the movement of each planar object (i.e., historical H). To be specific, we factorize the sensitive homography matrix, which describes the relative movement of planar objects, into parameters of transition, rotation, and pose. Finally, with the prediction from the mentioned stages, we propose a two-stream self-attention network to jointly refine the predicted planar objects."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 304,
+ 316,
+ 547,
+ 388
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 316,
+ 547,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 316,
+ 547,
+ 388
+ ],
+ "type": "text",
+ "content": "Besides, since there is no available dataset in this field, we build a large-scale benchmark dataset, namely MPOT-3K. Specifically, we shoot 356 videos with 3,717 planar objects and 687,417 annotations. The videos are collected under 9 motion patterns, where the relative movement and occlusion are also included to simulate the real-world scene."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 304,
+ 388,
+ 547,
+ 497
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 388,
+ 547,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 388,
+ 547,
+ 497
+ ],
+ "type": "text",
+ "content": "The contributions of this paper are two-fold: 1) We collect and annotate the first large-scale benchmark dataset for MPOT, where planar objects are diversely collected expelling bias. Our dataset will be released and boost research in this field. 2) We propose a tracking framework with unified motion and appearance models, which can accurately predict the pose and location of planar objects. Extensive experiments demonstrate the superiority of PRTrack against state-of-the-art approaches."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 306,
+ 507,
+ 392,
+ 520
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 507,
+ 392,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 507,
+ 392,
+ 520
+ ],
+ "type": "text",
+ "content": "2. Related Work"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 306,
+ 527,
+ 440,
+ 540
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 527,
+ 440,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 527,
+ 440,
+ 540
+ ],
+ "type": "text",
+ "content": "2.1. Planar Object Tracking"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 304,
+ 545,
+ 547,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 545,
+ 547,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 545,
+ 547,
+ 713
+ ],
+ "type": "text",
+ "content": "Benchmark Datasets. Current works focus on single planar object tracking and propose several POT datasets [19, 42, 43, 45, 70, 86]. Metaio [45] is the first dataset collected in the lab and the videos are collected by a monitored camera. Similarly, TMT [70] also shoots videos from the lab, where the annotations are automatically generated by aggregating the results of three planar object trackers. Due to the laboratory setting of the collection, videos in these datasets have similar backgrounds. Thus, POT210 [43] collects video data in the wild under seven motion patterns and encourages the diversity of the recorded videos with complex backgrounds. However, existing benchmarks are insufficient to mimic real-world settings where multiple targets exist. Therefore, our work extends the task into MPOT that"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 293,
+ 748,
+ 317,
+ 758
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 748,
+ 317,
+ 758
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 748,
+ 317,
+ 758
+ ],
+ "type": "text",
+ "content": "23461"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 49,
+ 92,
+ 286,
+ 163
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 47,
+ 71,
+ 287,
+ 92
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 71,
+ 287,
+ 92
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 71,
+ 287,
+ 92
+ ],
+ "type": "text",
+ "content": "Table 2: Statistics of tracking datasets. \"I.\" and \"O.\" indicate indoor and outdoor, respectively. \"#\" denotes \"the number of\"."
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 49,
+ 92,
+ 286,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 92,
+ 286,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 92,
+ 286,
+ 163
+ ],
+ "type": "table",
+ "html": "| Methods | Traditional | Deep | Ours |
| Motion Patterns | Metrics | CMT [61] | NCC [71] | CCRE [76] | MI [17] | GOP [8] | Gracker [79] | STM [62] | PoST [60] | LISRD [64] | SMask [78] | SRPN+ [38] | SPoint [16] | SOS [73] | GIFT [53] | HDN [95] |
| Overall | \\( S_{0.8} \\uparrow \\) | 09.50 | 00.19 | 09.99 | 26.70 | 56.03 | 67.92 | 09.07 | 33.52 | 63.43 | 64.04 | 67.23 | 71.82 | 72.91 | 75.86 | 76.40 | 82.51 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 64.27 | 61.95 | 34.34 | 37.82 | 62.28 | 88.38 | 52.05 | 39.55 | 72.39 | 73.29 | 75.37 | 71.17 | 70.06 | 72.99 | 79.50 | 94.59 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 37.05 | 03.64 | 14.12 | 23.57 | 45.25 | 85.67 | 25.64 | 25.02 | 56.23 | 58.38 | 60.70 | 54.58 | 53.23 | 56.41 | 65.21 | 92.85 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 10.24 | 00.55 | 19.08 | 38.60 | 62.72 | 78.25 | 23.07 | 40.74 | 80.40 | 69.55 | 74.45 | 81.25 | 85.37 | 84.02 | 82.56 | 90.78 |
| \\( D_{\\epsilon} \\downarrow \\) | 12.30 | 42.15 | 17.27 | 08.68 | 06.37 | 05.23 | 18.79 | 22.71 | 06.93 | 17.06 | 16.77 | 07.67 | 07.76 | 07.80 | 07.44 | 05.07 |
| Moving Occlusion | \\( S_{0.8} \\uparrow \\) | 11.34 | 00.00 | 10.69 | 26.47 | 55.85 | 63.68 | 08.78 | 18.17 | 73.79 | 63.29 | 70.62 | 70.91 | 73.42 | 72.92 | 74.96 | 80.29 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 67.25 | 62.78 | 56.21 | 68.10 | 81.41 | 92.93 | 55.49 | 48.04 | 91.74 | 81.46 | 83.27 | 89.92 | 92.28 | 89.74 | 88.74 | 93.16 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 53.48 | 4.92 | 33.56 | 51.82 | 70.05 | 76.30 | 29.63 | 25.45 | 86.37 | 74.57 | 75.94 | 84.73 | 84.35 | 84.21 | 82.00 | 90.95 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 13.47 | 00.64 | 32.01 | 61.57 | 77.30 | 81.98 | 24.37 | 28.98 | 80.70 | 67.62 | 72.97 | 85.39 | 85.03 | 85.50 | 84.86 | 88.27 |
| \\( D_{\\epsilon} \\downarrow \\) | 06.31 | 42.14 | 21.29 | 08.79 | 04.13 | 07.92 | 16.54 | 27.26 | 06.41 | 15.57 | 14.29 | 06.06 | 06.17 | 06.05 | 06.24 | 05.34 |
| Moving Objects | \\( S_{0.8} \\uparrow \\) | 09.34 | 00.69 | 11.23 | 24.29 | 52.21 | 61.58 | 08.09 | 35.52 | 67.79 | 62.95 | 65.49 | 65.50 | 70.36 | 67.02 | 70.09 | 79.12 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 69.22 | 63.86 | 40.78 | 47.41 | 68.96 | 92.54 | 53.77 | 41.96 | 76.88 | 80.33 | 82.72 | 78.44 | 69.29 | 78.59 | 81.05 | 94.29 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 60.28 | 17.10 | 19.24 | 32.57 | 52.50 | 90.48 | 28.33 | 27.13 | 60.87 | 68.62 | 72.09 | 67.75 | 52.41 | 67.97 | 67.18 | 92.27 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 22.87 | 2.19 | 24.27 | 53.95 | 72.51 | 56.77 | 25.28 | 43.96 | 87.73 | 75.75 | 79.22 | 67.66 | 86.09 | 67.79 | 84.43 | 90.44 |
| \\( D_{\\epsilon} \\downarrow \\) | 12.71 | 43.21 | 17.66 | 06.85 | 05.71 | 04.11 | 21.77 | 19.93 | 05.15 | 14.92 | 15.48 | 04.21 | 04.53 | 04.16 | 05.63 | 05.46 |
| Camera Occlusion | \\( S_{0.8} \\uparrow \\) | 14.61 | 00.00 | 13.31 | 41.79 | 63.59 | 66.41 | 13.80 | 40.21 | 85.46 | 62.46 | 59.40 | 88.85 | 88.28 | 89.43 | 84.01 | 92.46 |
| \\( Acc_{\\epsilon} \\uparrow \\) | 63.54 | 61.93 | 34.28 | 40.28 | 61.72 | 87.98 | 51.72 | 37.15 | 74.06 | 61.66 | 68.71 | 72.18 | 71.90 | 72.10 | 84.51 | 95.85 |
| \\( Pr_{\\epsilon} \\uparrow \\) | 36.89 | 00.38 | 18.42 | 27.84 | 45.04 | 88.24 | 24.59 | 25.31 | 56.76 | 44.83 | 52.31 | 54.67 | 54.42 | 54.60 | 73.16 | 95.84 |
| \\( Rc_{\\epsilon} \\uparrow \\) | 13.19 | 00.06 | 28.32 | 49.74 | 67.38 | 73.75 | 21.70 | 45.37 | 84.32 | 64.86 | 70.75 | 87.57 | 87.15 | 87.50 | 84.52 | 91.53 |
| \\( D_{\\epsilon} \\downarrow \\) | 10.31 | 29.38 | 20.10 | 07.71 | 06.73 | 04.63 | 25.05 | 20.08 | 04.82 | 14.46 | 14.45 | 04.81 | 04.62 | 04.73 | 05.18 | 04.59 |
| Speed↑ | 00.58 | 01.98 | 00.31 | 01.25 | 01.00 | 03.14 | 35.44 | 11.90 | 08.63 | 31.52 | 10.87 | 09.32 | 12.51 | 26.62 | 16.37 | 12.64 |
",
+ "image_path": "98bbb163c569c3f47ebff6bc0fbc7a304232b6d33a7b21c12bbb029e5f6b04f8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 47,
+ 346,
+ 288,
+ 369
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 346,
+ 288,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 346,
+ 288,
+ 369
+ ],
+ "type": "text",
+ "content": "ation. To fairly compare along with the trajectory level, we adopt the success rate "
+ },
+ {
+ "bbox": [
+ 47,
+ 346,
+ 288,
+ 369
+ ],
+ "type": "inline_equation",
+ "content": "(S_{\\alpha})"
+ },
+ {
+ "bbox": [
+ 47,
+ 346,
+ 288,
+ 369
+ ],
+ "type": "text",
+ "content": " with the distance threshold "
+ },
+ {
+ "bbox": [
+ 47,
+ 346,
+ 288,
+ 369
+ ],
+ "type": "inline_equation",
+ "content": "\\alpha"
+ },
+ {
+ "bbox": [
+ 47,
+ 346,
+ 288,
+ 369
+ ],
+ "type": "text",
+ "content": " as"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 69,
+ 378,
+ 287,
+ 409
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 69,
+ 378,
+ 287,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 69,
+ 378,
+ 287,
+ 409
+ ],
+ "type": "interline_equation",
+ "content": "S _ {\\alpha} = \\frac {| \\mathbb {C} |}{| \\mathbb {I} |}, \\mathbb {C} = \\left\\{i _ {n} \\in \\mathbb {I} \\right| \\frac {\\sum_ {t = 1} ^ {T} \\left| T _ {\\epsilon} ^ {t} \\right|}{\\sum_ {t = 1} ^ {T} \\left| G T _ {t} \\right|} > \\alpha \\}, \\tag {11}",
+ "image_path": "71a8face67308df86933d5f91a088cb8c681c0b0c3ad37987896f78a995e8a79.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 47,
+ 412,
+ 239,
+ 424
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 412,
+ 239,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 412,
+ 239,
+ 424
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 47,
+ 412,
+ 239,
+ 424
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbb{I}"
+ },
+ {
+ "bbox": [
+ 47,
+ 412,
+ 239,
+ 424
+ ],
+ "type": "text",
+ "content": " is the set of planar objects in the video."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 47,
+ 432,
+ 180,
+ 444
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 432,
+ 180,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 432,
+ 180,
+ 444
+ ],
+ "type": "text",
+ "content": "5.2. Implementation Details"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 47,
+ 451,
+ 287,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 451,
+ 287,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 451,
+ 287,
+ 713
+ ],
+ "type": "text",
+ "content": "We compare our model with fifteen representative methods. These methods are implemented by official code or open-source libraries [6, 72]. We run evaluated single object trackers multiple times for different planar objects. And we equip them with the data association strategy based on [12], for integrating each single object tracker into tracking multiple targets online. Besides, we additionally give the reference frames of ground truth if the single object trackers fail to track. Different from other box-based trackers, mask-based trackers (e.g., STM) directly outputs the mask. Thus, we use a rotated box estimator [88] to transfer the mask. We implement PRTrack using two RTX3090 GPUs. In the training stage, we use data augmentations, including flip, pepper noise, rotation, contrast jittering, and perspective transformation. We adopt a data sampling strategy following [41]. We adopt Smooth L1 loss and Cross-Entropy loss for optimizing the predicted vertex heatmaps and masks. Similar to [62], we pretrain SPB on YouTubeVOS [89] and DAVIS [66] datasets. During inference, we use the ordered vertexes to build the planar objects and estimate the homography matrix. We set the threshold as 0.9 in the reuse gate by grid search and set "
+ },
+ {
+ "bbox": [
+ 47,
+ 451,
+ 287,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "r"
+ },
+ {
+ "bbox": [
+ 47,
+ 451,
+ 287,
+ 713
+ ],
+ "type": "text",
+ "content": " in Equ. 2 as 0 during inference."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 306,
+ 345,
+ 429,
+ 357
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 345,
+ 429,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 345,
+ 429,
+ 357
+ ],
+ "type": "text",
+ "content": "5.3. Results on MPOT-3K"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 304,
+ 365,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 365,
+ 545,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 365,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "content": "We compare our PRTrack with the other fifteen trackers on MPOT-3K, which contains six traditional POT trackers, four deep-based generic object trackers, and five deep-based planar object trackers. Tab. 3 reports the overall performance as well as the one in the challenging situations of moving occlusion, moving objects, and camera occlusion."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "text",
+ "content": "Deep vs. Traditional Trackers. We can observe that the deep-based trackers have increased by "
+ },
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "inline_equation",
+ "content": "82.6\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "text",
+ "content": " on average, on five metrics for overall performance. This is because deep-based trackers avoid taking some assumptions under the laboratory environment, such as the invariance to illumination and appearance change. Note that Gracker performs best in traditional trackers and achieves competitive performance with the deep-based tracker group. However, since MPOT-3K has many small planar objects, it can not extract sufficient information, thus resulting in suboptimal performance (i.e., "
+ },
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "inline_equation",
+ "content": "10.3\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "text",
+ "content": " degradation in comparison with ours). In terms of deep-based trackers, STM predicts segmentation masks at the pixel level while ignoring the order of vertexes. Thus, it results in inferior performance, where only "
+ },
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "inline_equation",
+ "content": "9.0\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 437,
+ 545,
+ 616
+ ],
+ "type": "text",
+ "content": " of targets are correctly tracked."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 304,
+ 617,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 617,
+ 545,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 617,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": "Planar Object Trackers vs. Generic Object Trackers. Planar object trackers achieve higher performance (58.9% improvement) against generic object trackers, including SRPN+, SMask, STM, and PoST. A predominant reason is that the generic trackers are subject to the assumption of affine transformation, where only the coarse box estimation is required. Therefore, a powerful planar object tracker should be developed in this field."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 293,
+ 749,
+ 318,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 749,
+ 318,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 749,
+ 318,
+ 757
+ ],
+ "type": "text",
+ "content": "23466"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 48,
+ 45,
+ 166,
+ 145
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 48,
+ 45,
+ 166,
+ 145
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 45,
+ 166,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 45,
+ 166,
+ 145
+ ],
+ "type": "image",
+ "image_path": "9f635b7d3f32b0622c8b4a686a297d059450fb855b97c6f5a5256412428d9ae6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 167,
+ 45,
+ 287,
+ 146
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 167,
+ 45,
+ 287,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 45,
+ 287,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 45,
+ 287,
+ 146
+ ],
+ "type": "image",
+ "image_path": "5a567515146fced29c4d3cd7972f5a77d3892cde57c2ae027f18846aa3631cd2.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 89,
+ 147,
+ 241,
+ 243
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 85,
+ 146,
+ 129,
+ 156
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 85,
+ 146,
+ 129,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 85,
+ 146,
+ 129,
+ 156
+ ],
+ "type": "text",
+ "content": "(a) POT210"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 205,
+ 146,
+ 250,
+ 156
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 205,
+ 146,
+ 250,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 205,
+ 146,
+ 250,
+ 156
+ ],
+ "type": "text",
+ "content": "(b) POT280"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 89,
+ 147,
+ 241,
+ 243
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 89,
+ 147,
+ 241,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 89,
+ 147,
+ 241,
+ 243
+ ],
+ "type": "image",
+ "image_path": "2826847f90906e62103655908e4f5cf24e3dca71bebb4f4b1a88913615ef853e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 135,
+ 245,
+ 220,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 245,
+ 220,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 245,
+ 220,
+ 255
+ ],
+ "type": "text",
+ "content": "(c) Challenging factors"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 46,
+ 256,
+ 288,
+ 299
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 256,
+ 288,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 256,
+ 288,
+ 299
+ ],
+ "type": "text",
+ "content": "Figure 6: Comparison on POT210 and POT280. The averaged Precision on two datasets with different thresholds are reported in (a) and (b), respectively. (c) further shows the performance under seven challenging factors. The data is extracted from [95]."
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 46,
+ 312,
+ 287,
+ 396
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 312,
+ 287,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 312,
+ 287,
+ 396
+ ],
+ "type": "text",
+ "content": "PRTrack vs. SOTA Trackers. PRTrack outperforms other methods with an average improvement of "
+ },
+ {
+ "bbox": [
+ 46,
+ 312,
+ 287,
+ 396
+ ],
+ "type": "inline_equation",
+ "content": "6.4\\%"
+ },
+ {
+ "bbox": [
+ 46,
+ 312,
+ 287,
+ 396
+ ],
+ "type": "text",
+ "content": ". We also observe a large gain when there exists relative movement among multiple targets. To be specific, when facing challenging situations like moving objects, moving occlusion, and camera occlusion, PRTrack achieves a performance gain of "
+ },
+ {
+ "bbox": [
+ 46,
+ 312,
+ 287,
+ 396
+ ],
+ "type": "inline_equation",
+ "content": "16.9\\%"
+ },
+ {
+ "bbox": [
+ 46,
+ 312,
+ 287,
+ 396
+ ],
+ "type": "text",
+ "content": " against other advanced algorithms."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 47,
+ 403,
+ 222,
+ 415
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 403,
+ 222,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 403,
+ 222,
+ 415
+ ],
+ "type": "text",
+ "content": "5.4. Results on POT210 and POT280"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 46,
+ 422,
+ 287,
+ 578
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 422,
+ 287,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 422,
+ 287,
+ 578
+ ],
+ "type": "text",
+ "content": "To verify the effectiveness of PRTrack on the traditional POT task, we conduct experiments on the POT210 and POT280 datasets, and compare PRTrack with eleven POT methods as shown in Fig. 6. Note that we do not fine-tune PRTrack on these two datasets, again verifying the generalization ability of PRTrack towards unseen scenes. Following [95], the averaged Precision is employed to evaluate all the methods. As the distance threshold "
+ },
+ {
+ "bbox": [
+ 46,
+ 422,
+ 287,
+ 578
+ ],
+ "type": "inline_equation",
+ "content": "E_{AL}"
+ },
+ {
+ "bbox": [
+ 46,
+ 422,
+ 287,
+ 578
+ ],
+ "type": "text",
+ "content": " increases, the precision tends to be stable when "
+ },
+ {
+ "bbox": [
+ 46,
+ 422,
+ 287,
+ 578
+ ],
+ "type": "inline_equation",
+ "content": "E_{AL} > 10"
+ },
+ {
+ "bbox": [
+ 46,
+ 422,
+ 287,
+ 578
+ ],
+ "type": "text",
+ "content": ". We notice that PRTrack achieves competitive performance in comparison with advanced algorithms. Besides, PRTrack shows its superiority when handling challenging cases such as occlusion and motion blur."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 47,
+ 586,
+ 141,
+ 600
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 586,
+ 141,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 586,
+ 141,
+ 600
+ ],
+ "type": "text",
+ "content": "5.5. Ablation Study"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 287,
+ 676
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 605,
+ 287,
+ 676
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 605,
+ 287,
+ 676
+ ],
+ "type": "text",
+ "content": "We conduct ablation studies on MPOT-3K in Tab. 4 to explore the effectiveness of each component in PRTrack. Our baseline is the appearance perception module, without shifted encoder and auxiliary mask. For simplicity, we use \"i&j\" to denote the comparison between the "
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 287,
+ 676
+ ],
+ "type": "inline_equation",
+ "content": "i^{\\text{th}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 287,
+ 676
+ ],
+ "type": "text",
+ "content": " line and the "
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 287,
+ 676
+ ],
+ "type": "inline_equation",
+ "content": "j^{\\text{th}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 605,
+ 287,
+ 676
+ ],
+ "type": "text",
+ "content": " line in the rest of this subsection."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "text",
+ "content": "We first perform module-wise ablation studies in the highlighted lines, we observe that PRTrack obtains the largest performance gain of "
+ },
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "11.6\\%"
+ },
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "text",
+ "content": " (i.e., "
+ },
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "3 \\& 9"
+ },
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "4 \\& 10"
+ },
+ {
+ "bbox": [
+ 46,
+ 677,
+ 287,
+ 714
+ ],
+ "type": "text",
+ "content": ")"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "text",
+ "content": "Table 4: Ablation study of components and their variants of PRTrack on the validation split of MPOT-3K. The lines with background indicate the module-level ablation study. Here AP, OR, and RT are short of appearance perception, occlusion reasoning, and memory pool. SE and Fac denote the shifted encoder and factorized homography matrix. Msk represents the plane structure branch with mask information. Here, "
+ },
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "inline_equation",
+ "content": "*"
+ },
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "inline_equation",
+ "content": "\\bullet"
+ },
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "text",
+ "content": ", and "
+ },
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "inline_equation",
+ "content": "\\diamond"
+ },
+ {
+ "bbox": [
+ 305,
+ 46,
+ 545,
+ 140
+ ],
+ "type": "text",
+ "content": " are the fusion strategy (Fus), representing concatenation, attention-based, and indicator-based. Gat indicates the reuse gate."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 308,
+ 141,
+ 544,
+ 255
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 308,
+ 141,
+ 544,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 308,
+ 141,
+ 544,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 308,
+ 141,
+ 544,
+ 255
+ ],
+ "type": "table",
+ "html": "