| Method | CIFAR-10 | CIFAR-100 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 75.25±0.78 | 60.80±2.20 | 59.62±0.31 | 64.24±1.97 | 52.08±0.83 | 41.55±0.85 | 38.22±0.50 | 42.52±0.58 |
| ER (Rolnick et al., 2019) | 75.94±0.86 | 63.56±1.32 | 60.13±0.56 | 64.81±2.70 | 52.95±1.25 | 42.82±0.05 | 41.12±0.56 | 42.74±1.09 |
| ER-MIR (Aljundi et al., 2019) | 75.89±1.02 | 61.93±0.93 | 60.39±0.48 | 61.64±3.86 | 52.93±1.44 | 42.47±0.13 | 41.19±0.63 | 42.93±1.18 |
| REMIND (Hayes et al., 2020) | 69.55±0.91 | 53.34±1.01 | 58.01±0.72 | 59.27±1.86 | 40.87±0.76 | 36.17±1.83 | 23.40±2.25 | 28.78±1.71 |
| DER++ (Buzzega et al., 2020) | 74.78±0.72 | 59.20±0.95 | 59.44±0.28 | 66.11±1.80 | 38.16±1.57 | 38.55±2.11 | 29.38±2.58 | 38.20±3.13 |
| SCR (Mai et al., 2021) | 75.61±0.93 | 56.52±0.52 | 60.62±0.43 | 58.41±2.39 | 41.84±0.74 | 36.00±0.83 | 31.33±0.41 | 32.11±0.39 |
| ODDL (Ye et al., 2022) | 75.03±1.00 | 61.61±3.55 | 65.46±0.46 | 66.19±2.08 | 40.26±0.50 | 41.88±4.52 | 38.82±0.49 | 41.35±1.08 |
| MEMO (Zhou et al., 2023) | 73.21±0.49 | 62.47±3.38 | 59.26±0.90 | 62.01±1.17 | 40.60±1.11 | 39.87±0.46 | 23.41±1.63 | 32.74±2.11 |
| X-DER (Boschini et al., 2023) | 77.59±0.62 | 65.40±4.79 | 58.28±1.17 | 59.76±4.28 | 52.80±1.61 | 43.73±0.86 | 41.94±0.57 | 46.00±1.04 |
| EARL (Ours) | 78.31±0.72 | 65.71±2.26 | 69.52±0.19 | 70.41±1.97 | 57.12±1.22 | 44.40±0.68 | 47.89±0.61 | 46.09±0.26 |
",
+ "bbox": [
+ 78,
+ 88,
+ 893,
+ 277
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/7f31bf504023dff07cdaf10d49311d56658ca7aa529299a96d29a5e2cedf8e32.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Method | TinyImageNet | ImageNet-200 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 37.95±0.93 | 27.50±0.80 | 25.29±0.81 | 26.06±0.52 | 41.84±0.64 | 31.57±0.80 | 30.71±0.27 | 33.33±0.98 |
| ER (Rolnick et al., 2019) | 37.43±1.05 | 27.47±0.63 | 26.37±0.89 | 25.79±0.44 | 41.51±0.76 | 30.87±0.72 | 32.39±0.36 | 33.09±0.37 |
| ER-MIR (Aljundi et al., 2019) | 37.81±1.06 | 26.72±0.86 | 26.22±0.69 | 25.11±1.04 | 38.28±0.38 | 33.12±0.73 | 32.17±0.44 | 33.85±0.93 |
| REMIND (Hayes et al., 2020) | 28.37±0.13 | 27.68±0.45 | 10.19±0.60 | 14.90±1.49 | 39.25±0.93 | 31.98±0.84 | 30.23±0.62 | 33.98±0.09 |
| DER++ (Buzzega et al., 2020) | 39.38±0.60 | 29.36±0.81 | 27.23±1.94 | 31.53±0.80 | 43.50±0.31 | 34.56±0.50 | 35.22±0.26 | 38.38±0.97 |
| SCR (Mai et al., 2021) | 34.65±1.08 | 22.18±0.32 | 25.86±0.94 | 22.54±0.59 | 41.90±0.40 | 28.92±0.40 | 33.24±0.32 | 30.98±0.28 |
| MEMO (Zhou et al., 2023) | 27.36±0.61 | 27.57±0.52 | 10.82±1.23 | 18.03±1.36 | 41.55±0.23 | 34.19±1.47 | 32.54±0.39 | 36.11±1.06 |
| X-DER (Boschini et al., 2023) | 35.15±2.12 | 26.67±0.52 | 29.71±0.86 | 28.10±0.50 | 43.41±0.47 | 34.14±0.98 | 36.31±0.17 | 38.61±0.55 |
| EARL (Ours) | 41.77±1.26 | 29.65±0.20 | 35.08±0.70 | 32.49±1.21 | 44.88±0.29 | 34.79±0.55 | 39.14±0.47 | 38.83±0.35 |
",
+ "bbox": [
+ 78,
+ 286,
+ 890,
+ 463
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Table 1. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for CIFAR-10, CIFAR-100, TinyImageNet and ImageNet-200.",
+ "bbox": [
+ 75,
+ 473,
+ 892,
+ 502
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "remaining discrepancy between the classifiers and the features at inference.",
+ "bbox": [
+ 75,
+ 508,
+ 468,
+ 539
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "To select which of the stored residuals to use during inference, we not only store the residual, but also $f(x)$ to choose the stored residual closest to $f(x_{\\mathrm{infer}})$ . Therefore, we retain 'feature-residual' pairs in a 'feature-residual memory' $M = \\{(\\hat{h}_i,\\mathbf{r}_i)\\}_{i = 1}^N$ , where $\\hat{h}_i = \\hat{f} (x_i)$ , $\\mathbf{r}_i = \\mathbf{w}_{y_i} - \\hat{f} (x_i)$ , where $N$ is the size of feature-residual memory, and $\\mathbf{w}_{y_i}$ is the classifier vector for class $y_i$ .",
+ "bbox": [
+ 75,
+ 542,
+ 468,
+ 651
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "During inference, we select the $k$ nearest neighbor $\\hat{h}$ 's, i.e., $\\{\\hat{h}_{n_1},\\hat{h}_{n_1},\\dots ,\\hat{h}_{n_k}\\}$ from $\\{\\hat{h}_i\\}_{i = 1}^N$ , since using only the nearest residual for correction may lead to incorrect inference predictions if a wrong residual is selected from a different class. Finally, following ResMem [57], we calculate the residual correction term $\\mathbf{r}_{\\mathrm{eval}}$ by a weighted average of the corresponding residuals $\\{\\mathbf{r}_{n_1},\\mathbf{r}_{n_2},\\dots ,\\mathbf{r}_{n_k}\\}$ , with weights $\\{s_1,s_2,\\ldots ,s_k\\}$ that are inversely proportional to the distance from $\\hat{f} (x_{\\mathrm{eval}})$ , as:",
+ "bbox": [
+ 75,
+ 654,
+ 468,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {r} _ {\\text {e v a l}} = \\sum_ {i = 1} ^ {k} s _ {i} \\mathbf {r} _ {i}, \\quad s _ {i} = \\frac {e ^ {- (\\hat {f} (x _ {\\text {e v a l}}) - \\hat {h} _ {i}) / \\tau}}{\\sum_ {j = 1} ^ {k} e ^ {- (\\hat {f} (x _ {\\text {e v a l}}) - \\hat {h} _ {j}) / \\tau}}, \\tag {5}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 101,
+ 811,
+ 468,
+ 854
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "where $\\tau$ is a temperature hyperparameter. We add the residual-correcting term $\\mathbf{r}_{\\mathrm{eval}}$ to the model output $\\hat{f} (x_{\\mathrm{eval}})$",
+ "bbox": [
+ 75,
+ 869,
+ 468,
+ 902
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "to obtain the corrected output $\\hat{f} (x_{\\mathrm{eval}})_{\\mathrm{corrected}}$ as:",
+ "bbox": [
+ 498,
+ 508,
+ 816,
+ 526
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {f} \\left(x _ {\\text {e v a l}}\\right) _ {\\text {c o r r e c t e d}} = \\hat {f} \\left(x _ {\\text {e v a l}}\\right) + \\mathbf {r} _ {\\text {e v a l}}. \\tag {6}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 583,
+ 537,
+ 890,
+ 558
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "5. Experiments",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 573,
+ 633,
+ 590
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "5.1. Experimental setup",
+ "text_level": 1,
+ "bbox": [
+ 498,
+ 599,
+ 687,
+ 616
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "We perform experiments on four datasets: CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. For all datasets, our experiments are conducted on both a disjoint setup [35] and a Gaussian scheduled setup [45, 51]. We report the average and standard deviation results in three different seeds, except ImageNet-1k due to computational resource constraints [2, 27].",
+ "bbox": [
+ 496,
+ 623,
+ 890,
+ 729
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "To evaluate anytime inference performance, we use the area under the curve accuracy $(A_{auc})$ [6, 27], which measures the area under the accuracy curve. We also use last accuracy $(A_{last})$ which measures accuracy at the end of training. For detailed information about the experimental setup, refer to the Supplementary.",
+ "bbox": [
+ 496,
+ 731,
+ 890,
+ 821
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Baselines. We compare EARL with various CL methods in different categories: regularization (EWC [26]), network expansion (MEMO [63]), replay (ER [43], ER-MIR [1], REMIND [20], SCR [31], ODDL [3]), and distillation",
+ "bbox": [
+ 496,
+ 839,
+ 892,
+ 900
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "23938",
+ "bbox": [
+ 478,
+ 944,
+ 519,
+ 957
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/4a92afd0f76460ee44d8922823af0b9d29866dc167005430069c9ea3c666b109.jpg",
+ "table_caption": [],
+ "table_footnote": [
+ "Table 2. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for ImageNet-1K."
+ ],
+ "table_body": "| Method | ImageNet-1K |
| Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 30.17 | 18.78 | 17.09 | 20.15 |
| ER (Rolnick et al., 2019) | 30.18 | 18.96 | 17.17 | 18.39 |
| ER-MIR (Aljundi et al., 2019) | 31.68 | 19.87 | 19.37 | 16.30 |
| REMIND (Hayes et al., 2020) | 32.47 | 24.59 | 17.42 | 19.79 |
| DER++ (Buzzega et al., 2020) | 31.08 | 18.98 | 23.73 | 23.14 |
| SCR (Mai et al., 2021) | 26.72 | 13.60 | 19.82 | 15.84 |
| MEMO (Zhou et al., 2023) | 30.45 | 19.08 | 14.06 | 20.12 |
| X-DER (Boschini et al., 2023) | 31.67 | 18.18 | 25.18 | 12.51 |
| EARL (Ours) | 34.33 | 23.19 | 30.53 | 29.48 |
",
+ "bbox": [
+ 80,
+ 89,
+ 467,
+ 253
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "$(\\mathrm{DER} + + [4], \\mathrm{X - DER}[3])$ . For more details on the implementation of these methods and a comparison with NCFSCIL [56], which attempts to induce neural collapse in the context of few-shot class incremental learning, please refer to the supplementary material.",
+ "bbox": [
+ 75,
+ 291,
+ 467,
+ 367
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Implementation details. We use three components to architect our model: a backbone network $g(\\cdot)$ , a projection MLP $p(\\cdot)$ , and a fixed ETF classifier $\\mathbf{W}_{\\mathrm{ETF}}$ (i.e., our model $f$ can be defined as $f(x) = p \\circ g(x)$ ). For the projection layer, we attach an MLP projection layer $p_{\\theta_p}$ to the output of the backbone network $g_{\\theta_g}$ , where $\\theta_g$ and $\\theta_p$ denote the parameters of the backbone network and the projection layer, respectively (i.e., the model $f$ can be defined as $f(x) = p \\circ g(x)$ ), following [9, 37, 56]. For all methods, we use ResNet-18 [21] as the backbone network.",
+ "bbox": [
+ 75,
+ 382,
+ 467,
+ 532
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Following [27, 59], we employ memory-only training, where a random batch is selected from memory at each iteration. Furthermore, for episodic memory sampling, EARL uses the Greedy Balanced Sampling strategy [40]. We describe the details about hyperparameters and the pseudocode of EARL in Supplementary for the sake of space.",
+ "bbox": [
+ 75,
+ 532,
+ 467,
+ 623
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5.2. Results",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 633,
+ 169,
+ 648
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We first compare the accuracy of online continual learning methods, including EARL, and summarize the results in Table 1. As shown in the table, EARL outperforms other baselines on all benchmarks, both in disjoint and Gaussian-scheduled setups, except $A_{\\mathrm{last}}$ in ImageNet-1K disjoint setup. In particular, high $A_{\\mathrm{AUC}}$ suggests that EARL outperforms other methods for all the time that the data stream is provided to the model, which implies that it can be used for inference at anytime.",
+ "bbox": [
+ 75,
+ 657,
+ 467,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Furthermore, EARL, SCR, ER, MIR, DER, and ODDL do not use task boundary information during training, i.e., task-free, in contrast to EWC, X-DER, MEMO, and REMIND which use task-boundary information, i.e., task-aware. Nevertheless, EARL outperforms these methods even in the disjoint setup where utilizing task boundary information is advantageous due to abrupt distribution shifts",
+ "bbox": [
+ 75,
+ 794,
+ 467,
+ 898
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "at the task boundaries, except MEMO in ImageNet-1K disjoint. Since MEMO not only uses the task boundary information but also expands the network per task, this extra advantage of using a larger model is emphasized in largescale datasets such as ImageNet which require a large learning capacity of the model. Please refer to the supplementary material for details of the computational budget comparison among the baseline methods.",
+ "bbox": [
+ 496,
+ 90,
+ 890,
+ 210
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Memory budget. We consider not only the size of episodic memory, but also the size of the model, logits, and other components, following MEMO [63]. In the case of the memory budget for CIFAR-10 and the architecture of ResNet-18 (size of 42.6MB), EARL incurs an additional memory cost of 0.2MB to store feature-residual pairs. EWC requires an additional cost of 85.2MB due to the need to store Fisher Information per parameter and the previous model. MEMO, which expands the network per task, incurs an additional cost of 32MB per task. DER also requires an additional cost of 0.1MB to store logits during distillation.",
+ "bbox": [
+ 496,
+ 215,
+ 890,
+ 381
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "5.3. Ablation study",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 392,
+ 648,
+ 407
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We conduct an ablation study on the two components of EARL, preparatory data training and residual correction, and summarize the results in Table 3. Although both components contribute to performance improvement, preparatory data training shows a larger gain in performance. Training with preparatory data improves the baseline by $2.7\\% \\sim 3.3\\%$ in $A_{auc}$ and $2.1\\% \\sim 4.4\\%$ in $A_{last}$ . When combined with residual correction, we observe further gains across all metrics in both disjoint and Gaussian-scheduled setups.",
+ "bbox": [
+ 496,
+ 415,
+ 890,
+ 550
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/5ec652ff9bfca509aaf3d5de8627a380793ab48d2e58d806828c8db6872f8c63.jpg",
+ "image_caption": [
+ "Figure 5. Average similarity between features of the most recently added class's samples and the closest classifier vectors of the old classes (CIFAR-10, Gaussian Scheduled). Baseline is a vanilla ETF model trained only using episodic memory."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 545,
+ 563,
+ 823,
+ 726
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "For further analysis, Fig. 6 visualizes the cosine similarity between the output features of 50 randomly selected test set samples of novel class '4' and the classifier vectors $\\mathbf{w}_i$ . In baseline (a), the features of the class 4 are strongly biased towards the classifier vectors of the old classes (i.e., $\\mathbf{w}_0$ , $\\mathbf{w}_1$ , $\\mathbf{w}_2$ , $\\mathbf{w}_3$ ), rather than the correct classifier, $\\mathbf{w}_4$ . (c)",
+ "bbox": [
+ 496,
+ 809,
+ 890,
+ 900
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "23939",
+ "bbox": [
+ 478,
+ 945,
+ 517,
+ 955
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "table",
+ "img_path": "images/91716e7bf6d09db1e5161c74dcb538b16eeaa3438f4750be450e1a49d0242171.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Method | CIFAR-10 | CIFAR-100 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 75.25±0.78 | 60.80±2.20 | 59.62±0.31 | 64.24±1.97 | 52.08±0.83 | 41.55±0.85 | 38.22±0.50 | 42.52±0.58 |
| ER (Rolnick et al., 2019) | 75.94±0.86 | 63.56±1.32 | 60.13±0.56 | 64.81±2.70 | 52.95±1.25 | 42.82±0.05 | 41.12±0.56 | 42.74±1.09 |
| ER-MIR (Aljundi et al., 2019) | 75.89±1.02 | 61.93±0.93 | 60.39±0.48 | 61.64±3.86 | 52.93±1.44 | 42.47±0.13 | 41.19±0.63 | 42.93±1.18 |
| REMIND (Hayes et al., 2020) | 69.55±0.91 | 53.34±1.01 | 58.01±0.72 | 59.27±1.86 | 40.87±0.76 | 36.17±1.83 | 23.40±2.25 | 28.78±1.71 |
| DER++ (Buzzega et al., 2020) | 74.78±0.72 | 59.20±0.95 | 59.44±0.28 | 66.11±1.80 | 38.16±1.57 | 38.55±2.11 | 29.38±2.58 | 38.20±3.13 |
| SCR (Mai et al., 2021) | 75.61±0.93 | 56.52±0.52 | 60.62±0.43 | 58.41±2.39 | 41.84±0.74 | 36.00±0.83 | 31.33±0.41 | 32.11±0.39 |
| ODDL (Ye et al., 2022) | 75.03±1.00 | 61.61±3.55 | 65.46±0.46 | 66.19±2.08 | 40.26±0.50 | 41.88±4.52 | 38.82±0.49 | 41.35±1.08 |
| MEMO (Zhou et al., 2023) | 73.21±0.49 | 62.47±3.38 | 59.26±0.90 | 62.01±1.17 | 40.60±1.11 | 39.87±0.46 | 23.41±1.63 | 32.74±2.11 |
| X-DER (Boschini et al., 2023) | 77.59±0.62 | 65.40±4.79 | 58.28±1.17 | 59.76±4.28 | 52.80±1.61 | 43.73±0.86 | 41.94±0.57 | 46.00±1.04 |
| EARL (Ours) | 78.31±0.72 | 65.71±2.26 | 69.52±0.19 | 70.41±1.97 | 57.12±1.22 | 44.40±0.68 | 47.89±0.61 | 46.09±0.26 |
"
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.08,
+ 0.287,
+ 0.891,
+ 0.464
+ ],
+ "angle": 0,
+ "content": "| Method | TinyImageNet | ImageNet-200 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 37.95±0.93 | 27.50±0.80 | 25.29±0.81 | 26.06±0.52 | 41.84±0.64 | 31.57±0.80 | 30.71±0.27 | 33.33±0.98 |
| ER (Rolnick et al., 2019) | 37.43±1.05 | 27.47±0.63 | 26.37±0.89 | 25.79±0.44 | 41.51±0.76 | 30.87±0.72 | 32.39±0.36 | 33.09±0.37 |
| ER-MIR (Aljundi et al., 2019) | 37.81±1.06 | 26.72±0.86 | 26.22±0.69 | 25.11±1.04 | 38.28±0.38 | 33.12±0.73 | 32.17±0.44 | 33.85±0.93 |
| REMIND (Hayes et al., 2020) | 28.37±0.13 | 27.68±0.45 | 10.19±0.60 | 14.90±1.49 | 39.25±0.93 | 31.98±0.84 | 30.23±0.62 | 33.98±0.09 |
| DER++ (Buzzega et al., 2020) | 39.38±0.60 | 29.36±0.81 | 27.23±1.94 | 31.53±0.80 | 43.50±0.31 | 34.56±0.50 | 35.22±0.26 | 38.38±0.97 |
| SCR (Mai et al., 2021) | 34.65±1.08 | 22.18±0.32 | 25.86±0.94 | 22.54±0.59 | 41.90±0.40 | 28.92±0.40 | 33.24±0.32 | 30.98±0.28 |
| MEMO (Zhou et al., 2023) | 27.36±0.61 | 27.57±0.52 | 10.82±1.23 | 18.03±1.36 | 41.55±0.23 | 34.19±1.47 | 32.54±0.39 | 36.11±1.06 |
| X-DER (Boschini et al., 2023) | 35.15±2.12 | 26.67±0.52 | 29.71±0.86 | 28.10±0.50 | 43.41±0.47 | 34.14±0.98 | 36.31±0.17 | 38.61±0.55 |
| EARL (Ours) | 41.77±1.26 | 29.65±0.20 | 35.08±0.70 | 32.49±1.21 | 44.88±0.29 | 34.79±0.55 | 39.14±0.47 | 38.83±0.35 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.076,
+ 0.474,
+ 0.893,
+ 0.503
+ ],
+ "angle": 0,
+ "content": "Table 1. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for CIFAR-10, CIFAR-100, TinyImageNet and ImageNet-200."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.51,
+ 0.47,
+ 0.54
+ ],
+ "angle": 0,
+ "content": "remaining discrepancy between the classifiers and the features at inference."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.544,
+ 0.47,
+ 0.652
+ ],
+ "angle": 0,
+ "content": "To select which of the stored residuals to use during inference, we not only store the residual, but also \\( f(x) \\) to choose the stored residual closest to \\( f(x_{\\mathrm{infer}}) \\). Therefore, we retain 'feature-residual' pairs in a 'feature-residual memory' \\( M = \\{(\\hat{h}_i,\\mathbf{r}_i)\\}_{i = 1}^N \\), where \\( \\hat{h}_i = \\hat{f} (x_i) \\), \\( \\mathbf{r}_i = \\mathbf{w}_{y_i} - \\hat{f} (x_i) \\), where \\( N \\) is the size of feature-residual memory, and \\( \\mathbf{w}_{y_i} \\) is the classifier vector for class \\( y_i \\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.655,
+ 0.47,
+ 0.794
+ ],
+ "angle": 0,
+ "content": "During inference, we select the \\(k\\) nearest neighbor \\(\\hat{h}\\)'s, i.e., \\(\\{\\hat{h}_{n_1},\\hat{h}_{n_1},\\dots ,\\hat{h}_{n_k}\\}\\) from \\(\\{\\hat{h}_i\\}_{i = 1}^N\\), since using only the nearest residual for correction may lead to incorrect inference predictions if a wrong residual is selected from a different class. Finally, following ResMem [57], we calculate the residual correction term \\(\\mathbf{r}_{\\mathrm{eval}}\\) by a weighted average of the corresponding residuals \\(\\{\\mathbf{r}_{n_1},\\mathbf{r}_{n_2},\\dots ,\\mathbf{r}_{n_k}\\}\\), with weights \\(\\{s_1,s_2,\\ldots ,s_k\\}\\) that are inversely proportional to the distance from \\(\\hat{f} (x_{\\mathrm{eval}})\\), as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.102,
+ 0.812,
+ 0.47,
+ 0.856
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathbf {r} _ {\\text {e v a l}} = \\sum_ {i = 1} ^ {k} s _ {i} \\mathbf {r} _ {i}, \\quad s _ {i} = \\frac {e ^ {- (\\hat {f} (x _ {\\text {e v a l}}) - \\hat {h} _ {i}) / \\tau}}{\\sum_ {j = 1} ^ {k} e ^ {- (\\hat {f} (x _ {\\text {e v a l}}) - \\hat {h} _ {j}) / \\tau}}, \\tag {5}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.871,
+ 0.47,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "where \\(\\tau\\) is a temperature hyperparameter. We add the residual-correcting term \\(\\mathbf{r}_{\\mathrm{eval}}\\) to the model output \\(\\hat{f} (x_{\\mathrm{eval}})\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.499,
+ 0.509,
+ 0.818,
+ 0.527
+ ],
+ "angle": 0,
+ "content": "to obtain the corrected output \\(\\hat{f} (x_{\\mathrm{eval}})_{\\mathrm{corrected}}\\) as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.584,
+ 0.539,
+ 0.892,
+ 0.559
+ ],
+ "angle": 0,
+ "content": "\\[\n\\hat {f} \\left(x _ {\\text {e v a l}}\\right) _ {\\text {c o r r e c t e d}} = \\hat {f} \\left(x _ {\\text {e v a l}}\\right) + \\mathbf {r} _ {\\text {e v a l}}. \\tag {6}\n\\]"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.499,
+ 0.574,
+ 0.634,
+ 0.591
+ ],
+ "angle": 0,
+ "content": "5. Experiments"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.499,
+ 0.6,
+ 0.688,
+ 0.617
+ ],
+ "angle": 0,
+ "content": "5.1. Experimental setup"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.624,
+ 0.892,
+ 0.731
+ ],
+ "angle": 0,
+ "content": "We perform experiments on four datasets: CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. For all datasets, our experiments are conducted on both a disjoint setup [35] and a Gaussian scheduled setup [45, 51]. We report the average and standard deviation results in three different seeds, except ImageNet-1k due to computational resource constraints [2, 27]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.732,
+ 0.892,
+ 0.822
+ ],
+ "angle": 0,
+ "content": "To evaluate anytime inference performance, we use the area under the curve accuracy \\((A_{auc})\\) [6, 27], which measures the area under the accuracy curve. We also use last accuracy \\((A_{last})\\) which measures accuracy at the end of training. For detailed information about the experimental setup, refer to the Supplementary."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.84,
+ 0.893,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Baselines. We compare EARL with various CL methods in different categories: regularization (EWC [26]), network expansion (MEMO [63]), replay (ER [43], ER-MIR [1], REMIND [20], SCR [31], ODDL [3]), and distillation"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.479,
+ 0.945,
+ 0.521,
+ 0.958
+ ],
+ "angle": 0,
+ "content": "23938"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.081,
+ 0.09,
+ 0.468,
+ 0.255
+ ],
+ "angle": 0,
+ "content": "| Method | ImageNet-1K |
| Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 30.17 | 18.78 | 17.09 | 20.15 |
| ER (Rolnick et al., 2019) | 30.18 | 18.96 | 17.17 | 18.39 |
| ER-MIR (Aljundi et al., 2019) | 31.68 | 19.87 | 19.37 | 16.30 |
| REMIND (Hayes et al., 2020) | 32.47 | 24.59 | 17.42 | 19.79 |
| DER++ (Buzzega et al., 2020) | 31.08 | 18.98 | 23.73 | 23.14 |
| SCR (Mai et al., 2021) | 26.72 | 13.60 | 19.82 | 15.84 |
| MEMO (Zhou et al., 2023) | 30.45 | 19.08 | 14.06 | 20.12 |
| X-DER (Boschini et al., 2023) | 31.67 | 18.18 | 25.18 | 12.51 |
| EARL (Ours) | 34.33 | 23.19 | 30.53 | 29.48 |
"
+ },
+ {
+ "type": "table_footnote",
+ "bbox": [
+ 0.076,
+ 0.256,
+ 0.47,
+ 0.283
+ ],
+ "angle": 0,
+ "content": "Table 2. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for ImageNet-1K."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.292,
+ 0.468,
+ 0.368
+ ],
+ "angle": 0,
+ "content": "\\((\\mathrm{DER} + + [4], \\mathrm{X - DER}[3])\\). For more details on the implementation of these methods and a comparison with NCFSCIL [56], which attempts to induce neural collapse in the context of few-shot class incremental learning, please refer to the supplementary material."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.383,
+ 0.468,
+ 0.533
+ ],
+ "angle": 0,
+ "content": "Implementation details. We use three components to architect our model: a backbone network \\( g(\\cdot) \\), a projection MLP \\( p(\\cdot) \\), and a fixed ETF classifier \\( \\mathbf{W}_{\\mathrm{ETF}} \\) (i.e., our model \\( f \\) can be defined as \\( f(x) = p \\circ g(x) \\)). For the projection layer, we attach an MLP projection layer \\( p_{\\theta_p} \\) to the output of the backbone network \\( g_{\\theta_g} \\), where \\( \\theta_g \\) and \\( \\theta_p \\) denote the parameters of the backbone network and the projection layer, respectively (i.e., the model \\( f \\) can be defined as \\( f(x) = p \\circ g(x) \\)), following [9, 37, 56]. For all methods, we use ResNet-18 [21] as the backbone network."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.534,
+ 0.468,
+ 0.624
+ ],
+ "angle": 0,
+ "content": "Following [27, 59], we employ memory-only training, where a random batch is selected from memory at each iteration. Furthermore, for episodic memory sampling, EARL uses the Greedy Balanced Sampling strategy [40]. We describe the details about hyperparameters and the pseudocode of EARL in Supplementary for the sake of space."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.077,
+ 0.635,
+ 0.17,
+ 0.649
+ ],
+ "angle": 0,
+ "content": "5.2. Results"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.659,
+ 0.468,
+ 0.794
+ ],
+ "angle": 0,
+ "content": "We first compare the accuracy of online continual learning methods, including EARL, and summarize the results in Table 1. As shown in the table, EARL outperforms other baselines on all benchmarks, both in disjoint and Gaussian-scheduled setups, except \\( A_{\\mathrm{last}} \\) in ImageNet-1K disjoint setup. In particular, high \\( A_{\\mathrm{AUC}} \\) suggests that EARL outperforms other methods for all the time that the data stream is provided to the model, which implies that it can be used for inference at anytime."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.795,
+ 0.468,
+ 0.9
+ ],
+ "angle": 0,
+ "content": "Furthermore, EARL, SCR, ER, MIR, DER, and ODDL do not use task boundary information during training, i.e., task-free, in contrast to EWC, X-DER, MEMO, and REMIND which use task-boundary information, i.e., task-aware. Nevertheless, EARL outperforms these methods even in the disjoint setup where utilizing task boundary information is advantageous due to abrupt distribution shifts"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.092,
+ 0.892,
+ 0.212
+ ],
+ "angle": 0,
+ "content": "at the task boundaries, except MEMO in ImageNet-1K disjoint. Since MEMO not only uses the task boundary information but also expands the network per task, this extra advantage of using a larger model is emphasized in largescale datasets such as ImageNet which require a large learning capacity of the model. Please refer to the supplementary material for details of the computational budget comparison among the baseline methods."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.216,
+ 0.892,
+ 0.382
+ ],
+ "angle": 0,
+ "content": "Memory budget. We consider not only the size of episodic memory, but also the size of the model, logits, and other components, following MEMO [63]. In the case of the memory budget for CIFAR-10 and the architecture of ResNet-18 (size of 42.6MB), EARL incurs an additional memory cost of 0.2MB to store feature-residual pairs. EWC requires an additional cost of 85.2MB due to the need to store Fisher Information per parameter and the previous model. MEMO, which expands the network per task, incurs an additional cost of 32MB per task. DER also requires an additional cost of 0.1MB to store logits during distillation."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.393,
+ 0.65,
+ 0.408
+ ],
+ "angle": 0,
+ "content": "5.3. Ablation study"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.416,
+ 0.892,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "We conduct an ablation study on the two components of EARL, preparatory data training and residual correction, and summarize the results in Table 3. Although both components contribute to performance improvement, preparatory data training shows a larger gain in performance. Training with preparatory data improves the baseline by \\(2.7\\% \\sim 3.3\\%\\) in \\(A_{auc}\\) and \\(2.1\\% \\sim 4.4\\%\\) in \\(A_{last}\\). When combined with residual correction, we observe further gains across all metrics in both disjoint and Gaussian-scheduled setups."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.546,
+ 0.564,
+ 0.824,
+ 0.727
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.498,
+ 0.738,
+ 0.892,
+ 0.794
+ ],
+ "angle": 0,
+ "content": "Figure 5. Average similarity between features of the most recently added class's samples and the closest classifier vectors of the old classes (CIFAR-10, Gaussian Scheduled). Baseline is a vanilla ETF model trained only using episodic memory."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.81,
+ 0.892,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "For further analysis, Fig. 6 visualizes the cosine similarity between the output features of 50 randomly selected test set samples of novel class '4' and the classifier vectors \\(\\mathbf{w}_i\\). In baseline (a), the features of the class 4 are strongly biased towards the classifier vectors of the old classes (i.e., \\(\\mathbf{w}_0\\), \\(\\mathbf{w}_1\\), \\(\\mathbf{w}_2\\), \\(\\mathbf{w}_3\\)), rather than the correct classifier, \\(\\mathbf{w}_4\\). (c)"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.479,
+ 0.946,
+ 0.519,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "23939"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.084,
+ 0.077,
+ 0.887,
+ 0.179
+ ],
+ "angle": 0,
+ "content": "| Method | CIFAR-10 | CIFAR-100 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) | \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) | \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) | \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) |
| EWC (Kirkpatrick et al., 2017) | 75.25±0.78 | 60.80±2.20 | 59.62±0.31 | 64.24±1.97 | 52.08±0.83 | 41.55±0.85 | 38.22±0.50 | 42.52±0.58 |
| ER (Rolnick et al., 2019) | 75.94±0.86 | 63.56±1.32 | 60.13±0.56 | 64.81±2.70 | 52.95±1.25 | 42.82±0.05 | 41.12±0.56 | 42.74±1.09 |
| ER-MIR (Aljundi et al., 2019) | 75.89±1.02 | 61.93±0.93 | 60.39±0.48 | 61.64±3.86 | 52.93±1.44 | 42.47±0.13 | 41.19±0.63 | 42.93±1.18 |
| REMIND (Hayes et al., 2020) | 69.55±0.91 | 53.34±1.01 | 58.01±0.72 | 59.27±1.86 | 40.87±0.76 | 36.17±1.83 | 23.40±2.25 | 28.78±1.71 |
| DER++ (Buzzega et al., 2020) | 74.78±0.72 | 59.20±0.95 | 59.44±0.28 | 66.11±1.80 | 38.16±1.57 | 38.55±2.11 | 29.38±2.58 | 38.20±3.13 |
| SCR (Mai et al., 2021) | 75.61±0.93 | 56.52±0.52 | 60.62±0.43 | 58.41±2.39 | 41.84±0.74 | 36.00±0.83 | 31.33±0.41 | 32.11±0.39 |
| ODDL (Ye et al., 2022) | 75.03±1.00 | 61.61±3.55 | 65.46±0.46 | 66.19±2.08 | 40.26±0.50 | 41.88±4.52 | 38.82±0.49 | 41.35±1.08 |
| MEMO (Zhou et al., 2023) | 73.21±0.49 | 62.47±3.38 | 59.26±0.90 | 62.01±1.17 | 40.60±1.11 | 39.87±0.46 | 23.41±1.63 | 32.74±2.11 |
| X-DER (Boschini et al., 2023) | 77.59±0.62 | 65.40±4.79 | 58.28±1.17 | 59.76±4.28 | 52.80±1.61 | 43.73±0.86 | 41.94±0.57 | 46.00±1.04 |
| EARL (Ours) | 78.31±0.72 | 65.71±2.26 | 69.52±0.19 | 70.41±1.97 | 57.12±1.22 | 44.40±0.68 | 47.89±0.61 | 46.09±0.26 |
+
+| Method | TinyImageNet | ImageNet-200 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) | \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) | \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) | \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) |
| EWC (Kirkpatrick et al., 2017) | 37.95±0.93 | 27.50±0.80 | 25.29±0.81 | 26.06±0.52 | 41.84±0.64 | 31.57±0.80 | 30.71±0.27 | 33.33±0.98 |
| ER (Rolnick et al., 2019) | 37.43±1.05 | 27.47±0.63 | 26.37±0.89 | 25.79±0.44 | 41.51±0.76 | 30.87±0.72 | 32.39±0.36 | 33.09±0.37 |
| ER-MIR (Aljundi et al., 2019) | 37.81±1.06 | 26.72±0.86 | 26.22±0.69 | 25.11±1.04 | 38.28±0.38 | 33.12±0.73 | 32.17±0.44 | 33.85±0.93 |
| REMIND (Hayes et al., 2020) | 28.37±0.13 | 27.68±0.45 | 10.19±0.60 | 14.90±1.49 | 39.25±0.93 | 31.98±0.84 | 30.23±0.62 | 33.98±0.09 |
| DER++ (Buzzega et al., 2020) | 39.38±0.60 | 29.36±0.81 | 27.23±1.94 | 31.53±0.80 | 43.50±0.31 | 34.56±0.50 | 35.22±0.26 | 38.38±0.97 |
| SCR (Mai et al., 2021) | 34.65±1.08 | 22.18±0.32 | 25.86±0.94 | 22.54±0.59 | 41.90±0.40 | 28.92±0.40 | 33.24±0.32 | 30.98±0.28 |
| MEMO (Zhou et al., 2023) | 27.36±0.61 | 27.57±0.52 | 10.82±1.23 | 18.03±1.36 | 41.55±0.23 | 34.19±1.47 | 32.54±0.39 | 36.11±1.06 |
| X-DER (Boschini et al., 2023) | 35.15±2.12 | 26.67±0.52 | 29.71±0.86 | 28.10±0.50 | 43.41±0.47 | 34.14±0.98 | 36.31±0.17 | 38.61±0.55 |
| EARL (Ours) | 41.77±1.26 | 29.65±0.20 | 35.08±0.70 | 32.49±1.21 | 44.88±0.29 | 34.79±0.55 | 39.14±0.47 | 38.83±0.35 |
+
+Table 1. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for CIFAR-10, CIFAR-100, TinyImageNet and ImageNet-200.
+
+remaining discrepancy between the classifiers and the features at inference.
+
+To select which of the stored residuals to use during inference, we not only store the residual, but also $f(x)$ to choose the stored residual closest to $f(x_{\mathrm{infer}})$ . Therefore, we retain 'feature-residual' pairs in a 'feature-residual memory' $M = \{(\hat{h}_i,\mathbf{r}_i)\}_{i = 1}^N$ , where $\hat{h}_i = \hat{f} (x_i)$ , $\mathbf{r}_i = \mathbf{w}_{y_i} - \hat{f} (x_i)$ , where $N$ is the size of feature-residual memory, and $\mathbf{w}_{y_i}$ is the classifier vector for class $y_i$ .
+
+During inference, we select the $k$ nearest neighbor $\hat{h}$ 's, i.e., $\{\hat{h}_{n_1},\hat{h}_{n_1},\dots ,\hat{h}_{n_k}\}$ from $\{\hat{h}_i\}_{i = 1}^N$ , since using only the nearest residual for correction may lead to incorrect inference predictions if a wrong residual is selected from a different class. Finally, following ResMem [57], we calculate the residual correction term $\mathbf{r}_{\mathrm{eval}}$ by a weighted average of the corresponding residuals $\{\mathbf{r}_{n_1},\mathbf{r}_{n_2},\dots ,\mathbf{r}_{n_k}\}$ , with weights $\{s_1,s_2,\ldots ,s_k\}$ that are inversely proportional to the distance from $\hat{f} (x_{\mathrm{eval}})$ , as:
+
+$$
+\mathbf {r} _ {\text {e v a l}} = \sum_ {i = 1} ^ {k} s _ {i} \mathbf {r} _ {i}, \quad s _ {i} = \frac {e ^ {- (\hat {f} (x _ {\text {e v a l}}) - \hat {h} _ {i}) / \tau}}{\sum_ {j = 1} ^ {k} e ^ {- (\hat {f} (x _ {\text {e v a l}}) - \hat {h} _ {j}) / \tau}}, \tag {5}
+$$
+
+where $\tau$ is a temperature hyperparameter. We add the residual-correcting term $\mathbf{r}_{\mathrm{eval}}$ to the model output $\hat{f} (x_{\mathrm{eval}})$
+
+to obtain the corrected output $\hat{f} (x_{\mathrm{eval}})_{\mathrm{corrected}}$ as:
+
+$$
+\hat {f} \left(x _ {\text {e v a l}}\right) _ {\text {c o r r e c t e d}} = \hat {f} \left(x _ {\text {e v a l}}\right) + \mathbf {r} _ {\text {e v a l}}. \tag {6}
+$$
+
+# 5. Experiments
+
+# 5.1. Experimental setup
+
+We perform experiments on four datasets: CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. For all datasets, our experiments are conducted on both a disjoint setup [35] and a Gaussian scheduled setup [45, 51]. We report the average and standard deviation results in three different seeds, except ImageNet-1k due to computational resource constraints [2, 27].
+
+To evaluate anytime inference performance, we use the area under the curve accuracy $(A_{auc})$ [6, 27], which measures the area under the accuracy curve. We also use last accuracy $(A_{last})$ which measures accuracy at the end of training. For detailed information about the experimental setup, refer to the Supplementary.
+
+Baselines. We compare EARL with various CL methods in different categories: regularization (EWC [26]), network expansion (MEMO [63]), replay (ER [43], ER-MIR [1], REMIND [20], SCR [31], ODDL [3]), and distillation
+
+| Method | ImageNet-1K |
| Disjoint | Gaussian-Scheduled |
| \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) | \(A_{\text{AUC}} \uparrow\) | \(A_{\text{last}} \uparrow\) |
| EWC (Kirkpatrick et al., 2017) | 30.17 | 18.78 | 17.09 | 20.15 |
| ER (Rolnick et al., 2019) | 30.18 | 18.96 | 17.17 | 18.39 |
| ER-MIR (Aljundi et al., 2019) | 31.68 | 19.87 | 19.37 | 16.30 |
| REMIND (Hayes et al., 2020) | 32.47 | 24.59 | 17.42 | 19.79 |
| DER++ (Buzzega et al., 2020) | 31.08 | 18.98 | 23.73 | 23.14 |
| SCR (Mai et al., 2021) | 26.72 | 13.60 | 19.82 | 15.84 |
| MEMO (Zhou et al., 2023) | 30.45 | 19.08 | 14.06 | 20.12 |
| X-DER (Boschini et al., 2023) | 31.67 | 18.18 | 25.18 | 12.51 |
| EARL (Ours) | 34.33 | 23.19 | 30.53 | 29.48 |
+
+Table 2. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for ImageNet-1K.
+
+$(\mathrm{DER} + + [4], \mathrm{X - DER}[3])$ . For more details on the implementation of these methods and a comparison with NCFSCIL [56], which attempts to induce neural collapse in the context of few-shot class incremental learning, please refer to the supplementary material.
+
+Implementation details. We use three components to architect our model: a backbone network $g(\cdot)$ , a projection MLP $p(\cdot)$ , and a fixed ETF classifier $\mathbf{W}_{\mathrm{ETF}}$ (i.e., our model $f$ can be defined as $f(x) = p \circ g(x)$ ). For the projection layer, we attach an MLP projection layer $p_{\theta_p}$ to the output of the backbone network $g_{\theta_g}$ , where $\theta_g$ and $\theta_p$ denote the parameters of the backbone network and the projection layer, respectively (i.e., the model $f$ can be defined as $f(x) = p \circ g(x)$ ), following [9, 37, 56]. For all methods, we use ResNet-18 [21] as the backbone network.
+
+Following [27, 59], we employ memory-only training, where a random batch is selected from memory at each iteration. Furthermore, for episodic memory sampling, EARL uses the Greedy Balanced Sampling strategy [40]. We describe the details about hyperparameters and the pseudocode of EARL in Supplementary for the sake of space.
+
+# 5.2. Results
+
+We first compare the accuracy of online continual learning methods, including EARL, and summarize the results in Table 1. As shown in the table, EARL outperforms other baselines on all benchmarks, both in disjoint and Gaussian-scheduled setups, except $A_{\mathrm{last}}$ in ImageNet-1K disjoint setup. In particular, high $A_{\mathrm{AUC}}$ suggests that EARL outperforms other methods for all the time that the data stream is provided to the model, which implies that it can be used for inference at anytime.
+
+Furthermore, EARL, SCR, ER, MIR, DER, and ODDL do not use task boundary information during training, i.e., task-free, in contrast to EWC, X-DER, MEMO, and REMIND which use task-boundary information, i.e., task-aware. Nevertheless, EARL outperforms these methods even in the disjoint setup where utilizing task boundary information is advantageous due to abrupt distribution shifts
+
+at the task boundaries, except MEMO in ImageNet-1K disjoint. Since MEMO not only uses the task boundary information but also expands the network per task, this extra advantage of using a larger model is emphasized in largescale datasets such as ImageNet which require a large learning capacity of the model. Please refer to the supplementary material for details of the computational budget comparison among the baseline methods.
+
+Memory budget. We consider not only the size of episodic memory, but also the size of the model, logits, and other components, following MEMO [63]. In the case of the memory budget for CIFAR-10 and the architecture of ResNet-18 (size of 42.6MB), EARL incurs an additional memory cost of 0.2MB to store feature-residual pairs. EWC requires an additional cost of 85.2MB due to the need to store Fisher Information per parameter and the previous model. MEMO, which expands the network per task, incurs an additional cost of 32MB per task. DER also requires an additional cost of 0.1MB to store logits during distillation.
+
+# 5.3. Ablation study
+
+We conduct an ablation study on the two components of EARL, preparatory data training and residual correction, and summarize the results in Table 3. Although both components contribute to performance improvement, preparatory data training shows a larger gain in performance. Training with preparatory data improves the baseline by $2.7\% \sim 3.3\%$ in $A_{auc}$ and $2.1\% \sim 4.4\%$ in $A_{last}$ . When combined with residual correction, we observe further gains across all metrics in both disjoint and Gaussian-scheduled setups.
+
+
+Figure 5. Average similarity between features of the most recently added class's samples and the closest classifier vectors of the old classes (CIFAR-10, Gaussian Scheduled). Baseline is a vanilla ETF model trained only using episodic memory.
+
+For further analysis, Fig. 6 visualizes the cosine similarity between the output features of 50 randomly selected test set samples of novel class '4' and the classifier vectors $\mathbf{w}_i$ . In baseline (a), the features of the class 4 are strongly biased towards the classifier vectors of the old classes (i.e., $\mathbf{w}_0$ , $\mathbf{w}_1$ , $\mathbf{w}_2$ , $\mathbf{w}_3$ ), rather than the correct classifier, $\mathbf{w}_4$ . (c)
+
+| Method | CIFAR-10 | CIFAR-100 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 75.25±0.78 | 60.80±2.20 | 59.62±0.31 | 64.24±1.97 | 52.08±0.83 | 41.55±0.85 | 38.22±0.50 | 42.52±0.58 |
| ER (Rolnick et al., 2019) | 75.94±0.86 | 63.56±1.32 | 60.13±0.56 | 64.81±2.70 | 52.95±1.25 | 42.82±0.05 | 41.12±0.56 | 42.74±1.09 |
| ER-MIR (Aljundi et al., 2019) | 75.89±1.02 | 61.93±0.93 | 60.39±0.48 | 61.64±3.86 | 52.93±1.44 | 42.47±0.13 | 41.19±0.63 | 42.93±1.18 |
| REMIND (Hayes et al., 2020) | 69.55±0.91 | 53.34±1.01 | 58.01±0.72 | 59.27±1.86 | 40.87±0.76 | 36.17±1.83 | 23.40±2.25 | 28.78±1.71 |
| DER++ (Buzzega et al., 2020) | 74.78±0.72 | 59.20±0.95 | 59.44±0.28 | 66.11±1.80 | 38.16±1.57 | 38.55±2.11 | 29.38±2.58 | 38.20±3.13 |
| SCR (Mai et al., 2021) | 75.61±0.93 | 56.52±0.52 | 60.62±0.43 | 58.41±2.39 | 41.84±0.74 | 36.00±0.83 | 31.33±0.41 | 32.11±0.39 |
| ODDL (Ye et al., 2022) | 75.03±1.00 | 61.61±3.55 | 65.46±0.46 | 66.19±2.08 | 40.26±0.50 | 41.88±4.52 | 38.82±0.49 | 41.35±1.08 |
| MEMO (Zhou et al., 2023) | 73.21±0.49 | 62.47±3.38 | 59.26±0.90 | 62.01±1.17 | 40.60±1.11 | 39.87±0.46 | 23.41±1.63 | 32.74±2.11 |
| X-DER (Boschini et al., 2023) | 77.59±0.62 | 65.40±4.79 | 58.28±1.17 | 59.76±4.28 | 52.80±1.61 | 43.73±0.86 | 41.94±0.57 | 46.00±1.04 |
| EARL (Ours) | 78.31±0.72 | 65.71±2.26 | 69.52±0.19 | 70.41±1.97 | 57.12±1.22 | 44.40±0.68 | 47.89±0.61 | 46.09±0.26 |
",
+ "image_path": "a7c115f190f9210571617b661e4d4406938133cb087dd8120eb1d917fe9cebd5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 48,
+ 227,
+ 545,
+ 367
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 48,
+ 227,
+ 545,
+ 367
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 227,
+ 545,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 227,
+ 545,
+ 367
+ ],
+ "type": "table",
+ "html": "| Method | TinyImageNet | ImageNet-200 |
| Disjoint | Gaussian-Scheduled | Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 37.95±0.93 | 27.50±0.80 | 25.29±0.81 | 26.06±0.52 | 41.84±0.64 | 31.57±0.80 | 30.71±0.27 | 33.33±0.98 |
| ER (Rolnick et al., 2019) | 37.43±1.05 | 27.47±0.63 | 26.37±0.89 | 25.79±0.44 | 41.51±0.76 | 30.87±0.72 | 32.39±0.36 | 33.09±0.37 |
| ER-MIR (Aljundi et al., 2019) | 37.81±1.06 | 26.72±0.86 | 26.22±0.69 | 25.11±1.04 | 38.28±0.38 | 33.12±0.73 | 32.17±0.44 | 33.85±0.93 |
| REMIND (Hayes et al., 2020) | 28.37±0.13 | 27.68±0.45 | 10.19±0.60 | 14.90±1.49 | 39.25±0.93 | 31.98±0.84 | 30.23±0.62 | 33.98±0.09 |
| DER++ (Buzzega et al., 2020) | 39.38±0.60 | 29.36±0.81 | 27.23±1.94 | 31.53±0.80 | 43.50±0.31 | 34.56±0.50 | 35.22±0.26 | 38.38±0.97 |
| SCR (Mai et al., 2021) | 34.65±1.08 | 22.18±0.32 | 25.86±0.94 | 22.54±0.59 | 41.90±0.40 | 28.92±0.40 | 33.24±0.32 | 30.98±0.28 |
| MEMO (Zhou et al., 2023) | 27.36±0.61 | 27.57±0.52 | 10.82±1.23 | 18.03±1.36 | 41.55±0.23 | 34.19±1.47 | 32.54±0.39 | 36.11±1.06 |
| X-DER (Boschini et al., 2023) | 35.15±2.12 | 26.67±0.52 | 29.71±0.86 | 28.10±0.50 | 43.41±0.47 | 34.14±0.98 | 36.31±0.17 | 38.61±0.55 |
| EARL (Ours) | 41.77±1.26 | 29.65±0.20 | 35.08±0.70 | 32.49±1.21 | 44.88±0.29 | 34.79±0.55 | 39.14±0.47 | 38.83±0.35 |
",
+ "image_path": "7f31bf504023dff07cdaf10d49311d56658ca7aa529299a96d29a5e2cedf8e32.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 46,
+ 375,
+ 546,
+ 398
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 375,
+ 546,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 375,
+ 546,
+ 398
+ ],
+ "type": "text",
+ "content": "Table 1. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for CIFAR-10, CIFAR-100, TinyImageNet and ImageNet-200."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 46,
+ 403,
+ 287,
+ 427
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 403,
+ 287,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 403,
+ 287,
+ 427
+ ],
+ "type": "text",
+ "content": "remaining discrepancy between the classifiers and the features at inference."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": "To select which of the stored residuals to use during inference, we not only store the residual, but also "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "f(x)"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": " to choose the stored residual closest to "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "f(x_{\\mathrm{infer}})"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": ". Therefore, we retain 'feature-residual' pairs in a 'feature-residual memory' "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "M = \\{(\\hat{h}_i,\\mathbf{r}_i)\\}_{i = 1}^N"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{h}_i = \\hat{f} (x_i)"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{r}_i = \\mathbf{w}_{y_i} - \\hat{f} (x_i)"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "N"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": " is the size of feature-residual memory, and "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{w}_{y_i}"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": " is the classifier vector for class "
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "inline_equation",
+ "content": "y_i"
+ },
+ {
+ "bbox": [
+ 46,
+ 430,
+ 287,
+ 516
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": "During inference, we select the "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "k"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": " nearest neighbor "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{h}"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": "'s, i.e., "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "\\{\\hat{h}_{n_1},\\hat{h}_{n_1},\\dots ,\\hat{h}_{n_k}\\}"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": " from "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "\\{\\hat{h}_i\\}_{i = 1}^N"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": ", since using only the nearest residual for correction may lead to incorrect inference predictions if a wrong residual is selected from a different class. Finally, following ResMem [57], we calculate the residual correction term "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{r}_{\\mathrm{eval}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": " by a weighted average of the corresponding residuals "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "\\{\\mathbf{r}_{n_1},\\mathbf{r}_{n_2},\\dots ,\\mathbf{r}_{n_k}\\}"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": ", with weights "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "\\{s_1,s_2,\\ldots ,s_k\\}"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": " that are inversely proportional to the distance from "
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{f} (x_{\\mathrm{eval}})"
+ },
+ {
+ "bbox": [
+ 46,
+ 518,
+ 287,
+ 628
+ ],
+ "type": "text",
+ "content": ", as:"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 62,
+ 643,
+ 287,
+ 677
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 62,
+ 643,
+ 287,
+ 677
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 62,
+ 643,
+ 287,
+ 677
+ ],
+ "type": "interline_equation",
+ "content": "\\mathbf {r} _ {\\text {e v a l}} = \\sum_ {i = 1} ^ {k} s _ {i} \\mathbf {r} _ {i}, \\quad s _ {i} = \\frac {e ^ {- (\\hat {f} (x _ {\\text {e v a l}}) - \\hat {h} _ {i}) / \\tau}}{\\sum_ {j = 1} ^ {k} e ^ {- (\\hat {f} (x _ {\\text {e v a l}}) - \\hat {h} _ {j}) / \\tau}}, \\tag {5}",
+ "image_path": "04b39a20097d9a06aa53b7a8b4e19495f521cddab24e9ad0b5b5fc5a44284134.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\tau"
+ },
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "type": "text",
+ "content": " is a temperature hyperparameter. We add the residual-correcting term "
+ },
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{r}_{\\mathrm{eval}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "type": "text",
+ "content": " to the model output "
+ },
+ {
+ "bbox": [
+ 46,
+ 689,
+ 287,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{f} (x_{\\mathrm{eval}})"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 305,
+ 403,
+ 500,
+ 417
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 403,
+ 500,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 403,
+ 500,
+ 417
+ ],
+ "type": "text",
+ "content": "to obtain the corrected output "
+ },
+ {
+ "bbox": [
+ 305,
+ 403,
+ 500,
+ 417
+ ],
+ "type": "inline_equation",
+ "content": "\\hat{f} (x_{\\mathrm{eval}})_{\\mathrm{corrected}}"
+ },
+ {
+ "bbox": [
+ 305,
+ 403,
+ 500,
+ 417
+ ],
+ "type": "text",
+ "content": " as:"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 357,
+ 426,
+ 545,
+ 442
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 357,
+ 426,
+ 545,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 357,
+ 426,
+ 545,
+ 442
+ ],
+ "type": "interline_equation",
+ "content": "\\hat {f} \\left(x _ {\\text {e v a l}}\\right) _ {\\text {c o r r e c t e d}} = \\hat {f} \\left(x _ {\\text {e v a l}}\\right) + \\mathbf {r} _ {\\text {e v a l}}. \\tag {6}",
+ "image_path": "0930e031a2226f223906bf821288c360e6def32627b0c6fcc2f003f4c1daab12.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 305,
+ 454,
+ 388,
+ 468
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 454,
+ 388,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 454,
+ 388,
+ 468
+ ],
+ "type": "text",
+ "content": "5. Experiments"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 305,
+ 475,
+ 421,
+ 488
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 475,
+ 421,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 475,
+ 421,
+ 488
+ ],
+ "type": "text",
+ "content": "5.1. Experimental setup"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 304,
+ 494,
+ 545,
+ 578
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 494,
+ 545,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 494,
+ 545,
+ 578
+ ],
+ "type": "text",
+ "content": "We perform experiments on four datasets: CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-200, and ImageNet-1K. For all datasets, our experiments are conducted on both a disjoint setup [35] and a Gaussian scheduled setup [45, 51]. We report the average and standard deviation results in three different seeds, except ImageNet-1k due to computational resource constraints [2, 27]."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 304,
+ 579,
+ 545,
+ 651
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 579,
+ 545,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 579,
+ 545,
+ 651
+ ],
+ "type": "text",
+ "content": "To evaluate anytime inference performance, we use the area under the curve accuracy "
+ },
+ {
+ "bbox": [
+ 304,
+ 579,
+ 545,
+ 651
+ ],
+ "type": "inline_equation",
+ "content": "(A_{auc})"
+ },
+ {
+ "bbox": [
+ 304,
+ 579,
+ 545,
+ 651
+ ],
+ "type": "text",
+ "content": " [6, 27], which measures the area under the accuracy curve. We also use last accuracy "
+ },
+ {
+ "bbox": [
+ 304,
+ 579,
+ 545,
+ 651
+ ],
+ "type": "inline_equation",
+ "content": "(A_{last})"
+ },
+ {
+ "bbox": [
+ 304,
+ 579,
+ 545,
+ 651
+ ],
+ "type": "text",
+ "content": " which measures accuracy at the end of training. For detailed information about the experimental setup, refer to the Supplementary."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 304,
+ 665,
+ 546,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 665,
+ 546,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 665,
+ 546,
+ 713
+ ],
+ "type": "text",
+ "content": "Baselines. We compare EARL with various CL methods in different categories: regularization (EWC [26]), network expansion (MEMO [63]), replay (ER [43], ER-MIR [1], REMIND [20], SCR [31], ODDL [3]), and distillation"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 293,
+ 748,
+ 318,
+ 758
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 748,
+ 318,
+ 758
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 748,
+ 318,
+ 758
+ ],
+ "type": "text",
+ "content": "23938"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 49,
+ 71,
+ 286,
+ 201
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 49,
+ 71,
+ 286,
+ 201
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 71,
+ 286,
+ 201
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 71,
+ 286,
+ 201
+ ],
+ "type": "table",
+ "html": "| Method | ImageNet-1K |
| Disjoint | Gaussian-Scheduled |
| \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) | \\(A_{\\text{AUC}} \\uparrow\\) | \\(A_{\\text{last}} \\uparrow\\) |
| EWC (Kirkpatrick et al., 2017) | 30.17 | 18.78 | 17.09 | 20.15 |
| ER (Rolnick et al., 2019) | 30.18 | 18.96 | 17.17 | 18.39 |
| ER-MIR (Aljundi et al., 2019) | 31.68 | 19.87 | 19.37 | 16.30 |
| REMIND (Hayes et al., 2020) | 32.47 | 24.59 | 17.42 | 19.79 |
| DER++ (Buzzega et al., 2020) | 31.08 | 18.98 | 23.73 | 23.14 |
| SCR (Mai et al., 2021) | 26.72 | 13.60 | 19.82 | 15.84 |
| MEMO (Zhou et al., 2023) | 30.45 | 19.08 | 14.06 | 20.12 |
| X-DER (Boschini et al., 2023) | 31.67 | 18.18 | 25.18 | 12.51 |
| EARL (Ours) | 34.33 | 23.19 | 30.53 | 29.48 |
",
+ "image_path": "4a92afd0f76460ee44d8922823af0b9d29866dc167005430069c9ea3c666b109.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ },
+ {
+ "bbox": [
+ 46,
+ 202,
+ 287,
+ 224
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 202,
+ 287,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 202,
+ 287,
+ 224
+ ],
+ "type": "text",
+ "content": "Table 2. Comparison of online CL methods on Disjoint and Gaussian Scheduled Setup for ImageNet-1K."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_footnote"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 46,
+ 231,
+ 286,
+ 291
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 231,
+ 286,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 231,
+ 286,
+ 291
+ ],
+ "type": "inline_equation",
+ "content": "(\\mathrm{DER} + + [4], \\mathrm{X - DER}[3])"
+ },
+ {
+ "bbox": [
+ 46,
+ 231,
+ 286,
+ 291
+ ],
+ "type": "text",
+ "content": ". For more details on the implementation of these methods and a comparison with NCFSCIL [56], which attempts to induce neural collapse in the context of few-shot class incremental learning, please refer to the supplementary material."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": "Implementation details. We use three components to architect our model: a backbone network "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "g(\\cdot)"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": ", a projection MLP "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "p(\\cdot)"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": ", and a fixed ETF classifier "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{W}_{\\mathrm{ETF}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": " (i.e., our model "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "f"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": " can be defined as "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "f(x) = p \\circ g(x)"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": "). For the projection layer, we attach an MLP projection layer "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "p_{\\theta_p}"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": " to the output of the backbone network "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "g_{\\theta_g}"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": ", where "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "\\theta_g"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "\\theta_p"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": " denote the parameters of the backbone network and the projection layer, respectively (i.e., the model "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "f"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": " can be defined as "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "f(x) = p \\circ g(x)"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 286,
+ 422
+ ],
+ "type": "text",
+ "content": "), following [9, 37, 56]. For all methods, we use ResNet-18 [21] as the backbone network."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 46,
+ 422,
+ 286,
+ 494
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 422,
+ 286,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 422,
+ 286,
+ 494
+ ],
+ "type": "text",
+ "content": "Following [27, 59], we employ memory-only training, where a random batch is selected from memory at each iteration. Furthermore, for episodic memory sampling, EARL uses the Greedy Balanced Sampling strategy [40]. We describe the details about hyperparameters and the pseudocode of EARL in Supplementary for the sake of space."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 47,
+ 502,
+ 104,
+ 514
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 502,
+ 104,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 502,
+ 104,
+ 514
+ ],
+ "type": "text",
+ "content": "5.2. Results"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 46,
+ 521,
+ 286,
+ 628
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 521,
+ 286,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 521,
+ 286,
+ 628
+ ],
+ "type": "text",
+ "content": "We first compare the accuracy of online continual learning methods, including EARL, and summarize the results in Table 1. As shown in the table, EARL outperforms other baselines on all benchmarks, both in disjoint and Gaussian-scheduled setups, except "
+ },
+ {
+ "bbox": [
+ 46,
+ 521,
+ 286,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "A_{\\mathrm{last}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 521,
+ 286,
+ 628
+ ],
+ "type": "text",
+ "content": " in ImageNet-1K disjoint setup. In particular, high "
+ },
+ {
+ "bbox": [
+ 46,
+ 521,
+ 286,
+ 628
+ ],
+ "type": "inline_equation",
+ "content": "A_{\\mathrm{AUC}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 521,
+ 286,
+ 628
+ ],
+ "type": "text",
+ "content": " suggests that EARL outperforms other methods for all the time that the data stream is provided to the model, which implies that it can be used for inference at anytime."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 46,
+ 629,
+ 286,
+ 712
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 629,
+ 286,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 629,
+ 286,
+ 712
+ ],
+ "type": "text",
+ "content": "Furthermore, EARL, SCR, ER, MIR, DER, and ODDL do not use task boundary information during training, i.e., task-free, in contrast to EWC, X-DER, MEMO, and REMIND which use task-boundary information, i.e., task-aware. Nevertheless, EARL outperforms these methods even in the disjoint setup where utilizing task boundary information is advantageous due to abrupt distribution shifts"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 545,
+ 167
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 545,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 545,
+ 167
+ ],
+ "type": "text",
+ "content": "at the task boundaries, except MEMO in ImageNet-1K disjoint. Since MEMO not only uses the task boundary information but also expands the network per task, this extra advantage of using a larger model is emphasized in largescale datasets such as ImageNet which require a large learning capacity of the model. Please refer to the supplementary material for details of the computational budget comparison among the baseline methods."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 304,
+ 171,
+ 545,
+ 302
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 171,
+ 545,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 171,
+ 545,
+ 302
+ ],
+ "type": "text",
+ "content": "Memory budget. We consider not only the size of episodic memory, but also the size of the model, logits, and other components, following MEMO [63]. In the case of the memory budget for CIFAR-10 and the architecture of ResNet-18 (size of 42.6MB), EARL incurs an additional memory cost of 0.2MB to store feature-residual pairs. EWC requires an additional cost of 85.2MB due to the need to store Fisher Information per parameter and the previous model. MEMO, which expands the network per task, incurs an additional cost of 32MB per task. DER also requires an additional cost of 0.1MB to store logits during distillation."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 306,
+ 311,
+ 397,
+ 323
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 311,
+ 397,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 311,
+ 397,
+ 323
+ ],
+ "type": "text",
+ "content": "5.3. Ablation study"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "content": "We conduct an ablation study on the two components of EARL, preparatory data training and residual correction, and summarize the results in Table 3. Although both components contribute to performance improvement, preparatory data training shows a larger gain in performance. Training with preparatory data improves the baseline by "
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "inline_equation",
+ "content": "2.7\\% \\sim 3.3\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "content": " in "
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "inline_equation",
+ "content": "A_{auc}"
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "inline_equation",
+ "content": "2.1\\% \\sim 4.4\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "content": " in "
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "inline_equation",
+ "content": "A_{last}"
+ },
+ {
+ "bbox": [
+ 304,
+ 329,
+ 545,
+ 436
+ ],
+ "type": "text",
+ "content": ". When combined with residual correction, we observe further gains across all metrics in both disjoint and Gaussian-scheduled setups."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 334,
+ 446,
+ 504,
+ 575
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 334,
+ 446,
+ 504,
+ 575
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 334,
+ 446,
+ 504,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 334,
+ 446,
+ 504,
+ 575
+ ],
+ "type": "image",
+ "image_path": "5ec652ff9bfca509aaf3d5de8627a380793ab48d2e58d806828c8db6872f8c63.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 304,
+ 584,
+ 545,
+ 628
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 584,
+ 545,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 584,
+ 545,
+ 628
+ ],
+ "type": "text",
+ "content": "Figure 5. Average similarity between features of the most recently added class's samples and the closest classifier vectors of the old classes (CIFAR-10, Gaussian Scheduled). Baseline is a vanilla ETF model trained only using episodic memory."
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": "For further analysis, Fig. 6 visualizes the cosine similarity between the output features of 50 randomly selected test set samples of novel class '4' and the classifier vectors "
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{w}_i"
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": ". In baseline (a), the features of the class 4 are strongly biased towards the classifier vectors of the old classes (i.e., "
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{w}_0"
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{w}_1"
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{w}_2"
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": ", "
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{w}_3"
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": "), rather than the correct classifier, "
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{w}_4"
+ },
+ {
+ "bbox": [
+ 304,
+ 641,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": ". (c)"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 293,
+ 749,
+ 317,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 749,
+ 317,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 749,
+ 317,
+ 757
+ ],
+ "type": "text",
+ "content": "23939"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 51,
+ 60,
+ 542,
+ 141
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 542,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 542,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 60,
+ 542,
+ 141
+ ],
+ "type": "table",
+ "html": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| FCGF [14] + T-Linkage [33] | | 54.75 | 22.76 | 32.15 |
| FCGF [14] + RansaCov [34] | 36.27 | 73.50 | 45.01 | 55.83 |
| FCGF [14] + PointCLM [55] | | 86.99 | 70.05 | 77.60 |
| FCGF [14] + ECC [44] | | 92.60 | 73.79 | 82.13 |
| CoFiNet [54] + T-Linkage [33] | | 28.17 | 5.09 | 8.62 |
| CoFiNet [54] + RansaCov [34] | 18.52 | 38.29 | 13.69 | 20.16 |
| CoFiNet [54] + PointCLM [55] | | 36.15 | 24.70 | 29.34 |
| CoFiNet [54] + ECC [44] | | 57.26 | 17.24 | 26.50 |
| GeoTransformer [38] + T-Linkage [33] | | 72.68 | 44.80 | 55.43 |
| GeoTransformer [38] + RansaCov [34] | 44.02 | 78.84 | 67.16 | 72.53 |
| GeoTransformer [38] + PointCLM [55] | | 82.81 | 81.90 | 82.35 |
| GeoTransformer [38] + ECC [44] | | 94.63 | 74.83 | 83.57 |
| MIRETR (ours) + T-Linkage [33] | | 77.12 | 46.04 | 57.65 |
| MIRETR (ours) + RansaCov [34] | | 84.78 | 71.34 | 77.48 |
| MIRETR (ours) + PointCLM [55] | 56.59 | 91.85 | 91.08 | 91.46 |
| MIRETR (ours) + ECC [44] | | 96.52 | 89.03 | 92.62 |
| MIRETR (ours, full pipeline) | | 95.70 | 91.21 | 93.40 |
",
+ "bbox": [
+ 501,
+ 88,
+ 888,
+ 305
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Table 1. Evaluation results on Scan2CAD benchmark.",
+ "bbox": [
+ 532,
+ 309,
+ 854,
+ 321
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "5.1. Evaluations on Scan2CAD",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 330,
+ 741,
+ 345
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Dataset. Scan2CAD [2] is a scene-to-CAD alignment dataset build upon ScanNet [15] and ShapeNet [10]. It consists of 1506 scenes from ScanNet annotated with 14225 CAD models from ShapeNet and their poses in the scenes. Following [55], we replace the objects in the scenes with the corresponding aligned CAD models. We select the scene-model pairs which the scene contain multiple instances of the model for multi-instance registration. As last, we obtain 2184 point cloud pairs and use $70\\%$ pairs for training, $10\\%$ for validation and $20\\%$ for testing.",
+ "bbox": [
+ 496,
+ 354,
+ 890,
+ 503
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Quantitative results. As shown in Tab. 1, our MIRETR achieves consistent improvements over the baseline correspondence models. Notably, our method surpasses the previous best GeoTransformer on IR by 12 percentage points (pp). When coupled with PointCLM, it outperforms GeoTransformer by over 9 pp on the three registration metrics.",
+ "bbox": [
+ 496,
+ 505,
+ 890,
+ 595
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Moreover, our full pipeline surpasses all models with the baseline multi-model fitting methods. Although the ECC-based model performs slightly better than full MIRETR on MR, its MP is significantly worse. This indicates that ECC achieves high MR by predicting more registrations, which is impractical in real-world applications. And our method outperforms PointCLM on all three registration metrics, which demonstrates the strong effectiveness of our method.",
+ "bbox": [
+ 496,
+ 595,
+ 890,
+ 715
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "5.2. Evaluations on ROBI",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 726,
+ 700,
+ 739
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Dataset. ROBI [52] is a recent dataset for industrial bin-picking. It includes 7 reflective metallic industrial objects and 63 bin-picking scenes. Each scene contains cluttered instances of one industrial object. For each point cloud pair, the scene point cloud is backprojected from a depth image and the model point cloud is sampled from the CAD model of its corresponding industrial object. We obtain 4880 pairs in total, and we split the scenes for training, validation and testing according to the ratio of $6:1:2$ for each object.",
+ "bbox": [
+ 496,
+ 750,
+ 890,
+ 885
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Quantitative results. As in Tab. 2, our method outper",
+ "bbox": [
+ 500,
+ 886,
+ 890,
+ 901
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "19610",
+ "bbox": [
+ 480,
+ 944,
+ 519,
+ 955
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "image",
+ "img_path": "images/1abb0532efa0b1f690a72759bce9d97f471e82cd995a9373c97e80cf534597da.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 80,
+ 85,
+ 179,
+ 164
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/704b66b88c6729d9e6d53705b971c818475de3b48203da143327ec604af1982b.jpg",
+ "image_caption": [
+ "Inst:12"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 204,
+ 85,
+ 303,
+ 164
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/644149e1b7874c39a7f61a04f1fbf69b621909165d2e10ea50794c9885073d18.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 326,
+ 85,
+ 426,
+ 162
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/802cd1095a9446fc722eebc795ebfe0682251baab7a87e94de02204906d214f5.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 450,
+ 85,
+ 550,
+ 162
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/98d47ff7e66ebf5c4a5024041eec2b8facef842f7d79a9b042ef23194aeed501.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 563,
+ 85,
+ 718,
+ 162
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/dcf490464ac34f05aef2449727051e01a1176447869aa82803f6a3b80d6ebae8.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 730,
+ 85,
+ 885,
+ 164
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/69bf9dfd46a5d4c4b990e3f08d9f2ceea561adb39a0152fe649cafaaf2d4b944.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 80,
+ 178,
+ 179,
+ 255
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/a51edfd444ebd817b90e9e521858ae6e30c964e1c3b3c8b17636ec74df0ec497.jpg",
+ "image_caption": [
+ "Inst:22"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 202,
+ 178,
+ 303,
+ 255
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/61f2e6e3fb21f46a27a0863d5a0a1fe1e21d5f959140175d540a28ac1c232176.jpg",
+ "image_caption": [
+ "Inst:3"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 328,
+ 178,
+ 426,
+ 255
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/47222ca7df42dc2273d530c4ccfaf40b595aeb40c9ba23287a6c505bda959a24.jpg",
+ "image_caption": [
+ "Inst:9"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 450,
+ 178,
+ 550,
+ 255
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/9a7fcfb515a5db074b12478da6599137d1ac11a0ae2da99a77456ceb4d27c2a0.jpg",
+ "image_caption": [
+ "IR:39.42%",
+ "IR:49.91%"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 576,
+ 178,
+ 705,
+ 255
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/251cbcaee9afec1b852c631e6866a3942684bba4cd67373ccb205ee2290cc1e5.jpg",
+ "image_caption": [
+ "IR:54.21%",
+ "IR:56.90%"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 743,
+ 178,
+ 872,
+ 255
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/b69c08aeb868cc2fad1379dd2651a4718e7043dc9f9bed2dccc4cd7a2d810a69.jpg",
+ "image_caption": [
+ "(a) Input"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 78,
+ 271,
+ 181,
+ 351
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/88ba29007d8ff52a2871dd70323d5b0a20822679cfaf53c00c1570e020f5dcf1.jpg",
+ "image_caption": [
+ "Inst:18",
+ "(b) Ground Truth"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 202,
+ 272,
+ 303,
+ 351
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/d018345a684d4f108de169b40602c60da8d37b88e7b148d84d70253f3f89d44f.jpg",
+ "image_caption": [
+ "Inst:4",
+ "Inst:4",
+ "(c) GeoTrans"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 326,
+ 272,
+ 428,
+ 349
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/4e921912da9654eb7fc66703f378cf70ad42ea341c3022fb1f929244e3610718.jpg",
+ "image_caption": [
+ "Inst:11",
+ "Inst:8",
+ "(d) Ours"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 450,
+ 272,
+ 552,
+ 349
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/1c725f11c388f00e1a1af134e30d50ede5a2a17ff74c64aca2d5f3b1480fa8b4.jpg",
+ "image_caption": [
+ "IR:48.03%",
+ "(e) GeoTrans (corr)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 563,
+ 272,
+ 715,
+ 349
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/289dc0081fe5c9ddbfe53c9c7fbf6005030cc6947efdd95e0d5d9d14a572b8ed.jpg",
+ "image_caption": [
+ "IR:72.40%",
+ "(f) Ours (corr)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 730,
+ 272,
+ 883,
+ 349
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "table",
+ "img_path": "images/b12e836e86784d456a97df0e8d248d366226a281417fc839953ffa70b7175de9.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | 10.35 | 1.34 | 0.72 | 0.93 |
| CoFiNet [54] + RansaCov [34] | 1.25 | 1.39 | 1.31 |
| CoFiNet [54] + PointCLM [55] | 1.17 | 2.11 | 1.50 |
| CoFiNet [54] + ECC [44] | 3.21 | 8.50 | 4.66 |
| GeoTransformer [38] + T-Linkage [33] | 39.64 | 7.71 | 4.57 | 5.73 |
| GeoTransformer [38] + RansaCov [34] | 8.99 | 13.58 | 10.81 |
| GeoTransformer [38] + PointCLM [55] | 13.98 | 26.64 | 18.33 |
| GeoTransformer [38] + ECC [44] | 18.55 | 30.99 | 23.20 |
| MIRETR (ours) + T-Linkage [33] | 45.14 | 12.04 | 10.47 | 11.20 |
| MIRETR (ours) + RansaCov [34] | 14.14 | 26.29 | 18.38 |
| MIRETR (ours) + PointCLM [55] | 18.68 | 40.11 | 25.48 |
| MIRETR (ours) + ECC [44] | 24.65 | 34.85 | 28.91 |
| MIRETR (ours, full pipeline) | 38.51 | 41.19 | 39.80 |
",
+ "bbox": [
+ 78,
+ 433,
+ 465,
+ 602
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Table 2. Evaluation results on ROBI benchmark.",
+ "bbox": [
+ 127,
+ 607,
+ 416,
+ 619
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "forms the baseline correspondence methods by a large margin on all four metrics. Compared to Scan2CAD, ROBI is more challenging due to the more cluttered scenes, where instance-aware information is more important. Benefiting from the instance-aware geometric transformer module, our method extracts reliable instance-aware correspondences, which contributes to more accurate instance registrations.",
+ "bbox": [
+ 75,
+ 627,
+ 467,
+ 732
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Our full model achieves significant improvements over the ECC-based model on MR, MP and MF. The PointCLM-based model achieves comparable MP with the full model, but significantly worse MR, which means it can register only several easy instances but misses most heavily-occluded instances. Specifically, it can register on average only 6.2 instances per scene, while the number of our method is 13.7. Our method achieves both high MR and MP, showing strong robustness in cluttered scenes.",
+ "bbox": [
+ 75,
+ 734,
+ 467,
+ 869
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Qualitative results. Fig. 5 visualizes the registration results of GeoTransformer and MIRETR. Our method can",
+ "bbox": [
+ 75,
+ 869,
+ 467,
+ 898
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/391ca0191f955cb9a6ce937cf227c2f26566bf36dc2ed686e4e35c93a389f6c0.jpg",
+ "image_caption": [
+ "Figure 5. Registration results on ROBI benchmark. We visualize the successfully registered instances in (c) and (d). MIRETR registers more instances in the cluttered scenes (the $2^{\\text{nd}}$ and the $3^{\\text{rd}}$ rows) and with incomplete geometry (all three rows). And it extracts more accurate correspondences benefiting from the instance-aware correspondence learning mechanism."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 501,
+ 431,
+ 625,
+ 529
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/b6485a00d3b1d395f9bd361f2487dcd38fa3167e21cd81dbb7ffe8c613cfcf55.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 630,
+ 431,
+ 756,
+ 529
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/d5503b738ba1ee730cc57be49e2512510deae3912135aa643abd992b3e952ba8.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 761,
+ 431,
+ 888,
+ 529
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/8d3df5aa3f8921044baebe2827a67c205e342787a90d0bc6f09a89fe3001a9b0.jpg",
+ "image_caption": [
+ "(a) Anchor's neighbors",
+ "Figure 6. Visualizations of the predicted instance masks. The anchor superpoint (patch) is in red and the selected neighbors are color-coded. Our method can effectively reject the patches outside the interested instance while preserving most of those inside."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 501,
+ 531,
+ 627,
+ 628
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/d5e8662215981f9b8f3c9bdfbf1940a9049c4a266abfef58d8dc728417113bbf.jpg",
+ "image_caption": [
+ "(b) Ground Truth Mask"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 630,
+ 531,
+ 756,
+ 628
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/96232d5edff7c4e3fb917f7bd8549b8ca128637140de62e2cc227b322af338d4.jpg",
+ "image_caption": [
+ "(c) Our Prediceted Mask"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 761,
+ 531,
+ 888,
+ 628
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "register more instances in cluttered scenes (see the $2^{\\text{nd}}$ and the $3^{\\text{rd}}$ rows) and with severe geometric deficiency (see all three rows). And MIRETR attains more high-quality correspondences thanks to our instance-aware design.",
+ "bbox": [
+ 500,
+ 702,
+ 890,
+ 762
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We further visualize the predicted instance masks Fig. 6. The instance masks effectively reject the patches out of the interested instance while preserving most of those inside. Specifically, our method achieves the instance mIoU of 69.26 pp while the result of the model without masking is merely 38.13 pp. The high-quality masks allow us to effectively encode intra-instance context and contribute to accurate registrations.",
+ "bbox": [
+ 500,
+ 763,
+ 890,
+ 883
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Fig. 7 compares the results under different overlap ratios.",
+ "bbox": [
+ 517,
+ 885,
+ 888,
+ 900
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "19611",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/47fd1ca19d7110631edeade7b18fcf7f75922bb74cb3687da7de3747fb0f468c.jpg",
+ "image_caption": [
+ "Figure 7. Results in different overlap levels on ROBI benchmark."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 76,
+ 85,
+ 472,
+ 263
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "MIRETR consistently outperforms the baselines at all overlap levels, and the improvements are more significant when the overlap ratio is below $50\\%$ . This proves the strong capability of our method to handle low-overlap instances.",
+ "bbox": [
+ 76,
+ 287,
+ 468,
+ 349
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "5.3. Evaluations on ShapeNet",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 358,
+ 307,
+ 375
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Dataset. ShapeNet [10] is a large CAD model dataset which contains 51300 models from 55 categories. We use the models from first 30 categories for training and those from the rest 25 categories for testing to evaluate the generality to novel categories. We randomly sample at most 500 models from each category to avoid class imbalance. For each point cloud pair, the model point cloud is a random CAD model and the scene one is constructed by applying $4\\sim 16$ random poses on the model. At last, we obtain 8634 pairs for training, 900 for validation, and 7683 for testing.",
+ "bbox": [
+ 75,
+ 382,
+ 468,
+ 532
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Quantitative results. As shown in Tab. 3, our MIRETR achieves the best results on all four metrics, surpassing the baselines by a large margin. GeoTransformer significantly outperforms CoFiNet, which proves the importance of geometric structure to achieving generality. And our method achieves improvements of more than 10 pp over GeoTransformer on all registration metrics with PointCLM and ECC. Thanks to the instance-ware geometric transformer, the correspondences extracted by our method provide strong instance information, facilitating the correspondence clustering process. At last, our full model surpasses the previous best ECC-based baseline by around 2 pp on MR, 11 pp on MP and 7 pp on MF, demonstrating the strong generality of MIRETR to unseen objects.",
+ "bbox": [
+ 75,
+ 534,
+ 470,
+ 744
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "5.4. Ablation Studies",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 756,
+ 240,
+ 770
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We conduct extensive ablation studies on ROBI benchmark to better understand our design choices in Tab. 4. We first ablate the instance-aware point matching and extract point correspondences within only the patches of the superpoint correspondences in Tab. 4 (b), which is significantly worse than the full model. The poses estimated by this method is unstable due to clustered correspondences, thus the duplicated candidates cannot be effectively removed by NMS.",
+ "bbox": [
+ 75,
+ 779,
+ 468,
+ 902
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "table",
+ "img_path": "images/843107ec916c96881858bfb81362e9325f0099639ec1d5bd81622569773569d0.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | | 4.59 | 3.24 | 3.79 |
| CoFiNet [54] + RansaCov [34] | 39.53 | 7.35 | 5.33 | 6.17 |
| CoFiNet [54] + PointCLM [55] | | 23.40 | 19.66 | 21.36 |
| CoFiNet [54] + ECC [44] | | 41.49 | 26.60 | 32.41 |
| GeoTransformer [38] + T-Linkage [33] | | 27.68 | 26.40 | 27.02 |
| GeoTransformer [38] + RansaCov [34] | 69.65 | 41.86 | 48.16 | 44.78 |
| GeoTransformer [38] + PointCLM [55] | | 68.09 | 69.09 | 68.58 |
| GeoTransformer [38] + ECC [44] | | 78.03 | 64.08 | 70.37 |
| MIRETR (ours) + T-Linkage [33] | | 29.40 | 28.67 | 29.03 |
| MIRETR (ours) + RansaCov [34] | | 45.63 | 46.85 | 46.23 |
| MIRETR (ours) + PointCLM [55] | 77.13 | 85.63 | 85.81 | 85.71 |
| MIRETR (ours) + ECC [44] | | 92.56 | 82.08 | 87.01 |
| MIRETR (ours, full pipeline) | | 94.95 | 93.94 | 94.44 |
",
+ "bbox": [
+ 501,
+ 88,
+ 888,
+ 258
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "table",
+ "img_path": "images/1cff5e45d9ef139c4870cf53a0bfdeb7a6a006cc885ea66901caa584aa79197f.jpg",
+ "table_caption": [
+ "Table 3. Evaluation results on ShapeNet benchmark."
+ ],
+ "table_footnote": [],
+ "table_body": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| FCGF [14] + T-Linkage [33] | | 54.75 | 22.76 | 32.15 |
| FCGF [14] + RansaCov [34] | 36.27 | 73.50 | 45.01 | 55.83 |
| FCGF [14] + PointCLM [55] | | 86.99 | 70.05 | 77.60 |
| FCGF [14] + ECC [44] | | 92.60 | 73.79 | 82.13 |
| CoFiNet [54] + T-Linkage [33] | | 28.17 | 5.09 | 8.62 |
| CoFiNet [54] + RansaCov [34] | 18.52 | 38.29 | 13.69 | 20.16 |
| CoFiNet [54] + PointCLM [55] | | 36.15 | 24.70 | 29.34 |
| CoFiNet [54] + ECC [44] | | 57.26 | 17.24 | 26.50 |
| GeoTransformer [38] + T-Linkage [33] | | 72.68 | 44.80 | 55.43 |
| GeoTransformer [38] + RansaCov [34] | 44.02 | 78.84 | 67.16 | 72.53 |
| GeoTransformer [38] + PointCLM [55] | | 82.81 | 81.90 | 82.35 |
| GeoTransformer [38] + ECC [44] | | 94.63 | 74.83 | 83.57 |
| MIRETR (ours) + T-Linkage [33] | | 77.12 | 46.04 | 57.65 |
| MIRETR (ours) + RansaCov [34] | | 84.78 | 71.34 | 77.48 |
| MIRETR (ours) + PointCLM [55] | 56.59 | 91.85 | 91.08 | 91.46 |
| MIRETR (ours) + ECC [44] | | 96.52 | 89.03 | 92.62 |
| MIRETR (ours, full pipeline) | | 95.70 | 91.21 | 93.40 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.534,
+ 0.31,
+ 0.856,
+ 0.323
+ ],
+ "angle": 0,
+ "content": "Table 1. Evaluation results on Scan2CAD benchmark."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.332,
+ 0.742,
+ 0.346
+ ],
+ "angle": 0,
+ "content": "5.1. Evaluations on Scan2CAD"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.355,
+ 0.892,
+ 0.505
+ ],
+ "angle": 0,
+ "content": "Dataset. Scan2CAD [2] is a scene-to-CAD alignment dataset build upon ScanNet [15] and ShapeNet [10]. It consists of 1506 scenes from ScanNet annotated with 14225 CAD models from ShapeNet and their poses in the scenes. Following [55], we replace the objects in the scenes with the corresponding aligned CAD models. We select the scene-model pairs which the scene contain multiple instances of the model for multi-instance registration. As last, we obtain 2184 point cloud pairs and use \\(70\\%\\) pairs for training, \\(10\\%\\) for validation and \\(20\\%\\) for testing."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.506,
+ 0.892,
+ 0.597
+ ],
+ "angle": 0,
+ "content": "Quantitative results. As shown in Tab. 1, our MIRETR achieves consistent improvements over the baseline correspondence models. Notably, our method surpasses the previous best GeoTransformer on IR by 12 percentage points (pp). When coupled with PointCLM, it outperforms GeoTransformer by over 9 pp on the three registration metrics."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.597,
+ 0.892,
+ 0.717
+ ],
+ "angle": 0,
+ "content": "Moreover, our full pipeline surpasses all models with the baseline multi-model fitting methods. Although the ECC-based model performs slightly better than full MIRETR on MR, its MP is significantly worse. This indicates that ECC achieves high MR by predicting more registrations, which is impractical in real-world applications. And our method outperforms PointCLM on all three registration metrics, which demonstrates the strong effectiveness of our method."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.727,
+ 0.702,
+ 0.741
+ ],
+ "angle": 0,
+ "content": "5.2. Evaluations on ROBI"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.75,
+ 0.892,
+ 0.886
+ ],
+ "angle": 0,
+ "content": "Dataset. ROBI [52] is a recent dataset for industrial bin-picking. It includes 7 reflective metallic industrial objects and 63 bin-picking scenes. Each scene contains cluttered instances of one industrial object. For each point cloud pair, the scene point cloud is backprojected from a depth image and the model point cloud is sampled from the CAD model of its corresponding industrial object. We obtain 4880 pairs in total, and we split the scenes for training, validation and testing according to the ratio of \\( 6:1:2 \\) for each object."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.5,
+ 0.887,
+ 0.892,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "Quantitative results. As in Tab. 2, our method outper"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.52,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "19610"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.081,
+ 0.087,
+ 0.181,
+ 0.165
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.205,
+ 0.087,
+ 0.304,
+ 0.165
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.23,
+ 0.167,
+ 0.278,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "Inst:12"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.328,
+ 0.087,
+ 0.428,
+ 0.164
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.358,
+ 0.167,
+ 0.398,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "Inst:3"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.452,
+ 0.087,
+ 0.552,
+ 0.164
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.482,
+ 0.167,
+ 0.522,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "Inst:9"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.564,
+ 0.087,
+ 0.72,
+ 0.164
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.611,
+ 0.167,
+ 0.669,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "IR:39.42%"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.731,
+ 0.087,
+ 0.886,
+ 0.165
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.78,
+ 0.167,
+ 0.838,
+ 0.177
+ ],
+ "angle": 0,
+ "content": "IR:54.21%"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.081,
+ 0.179,
+ 0.18,
+ 0.256
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.204,
+ 0.179,
+ 0.304,
+ 0.256
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.23,
+ 0.259,
+ 0.278,
+ 0.269
+ ],
+ "angle": 0,
+ "content": "Inst:22"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.329,
+ 0.179,
+ 0.428,
+ 0.256
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.358,
+ 0.259,
+ 0.398,
+ 0.269
+ ],
+ "angle": 0,
+ "content": "Inst:4"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.452,
+ 0.179,
+ 0.552,
+ 0.256
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.479,
+ 0.259,
+ 0.524,
+ 0.27
+ ],
+ "angle": 0,
+ "content": "Inst:11"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.578,
+ 0.179,
+ 0.706,
+ 0.256
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.611,
+ 0.26,
+ 0.668,
+ 0.27
+ ],
+ "angle": 0,
+ "content": "IR:49.91%"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.745,
+ 0.179,
+ 0.874,
+ 0.256
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.78,
+ 0.26,
+ 0.838,
+ 0.27
+ ],
+ "angle": 0,
+ "content": "IR:56.90%"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.08,
+ 0.272,
+ 0.182,
+ 0.352
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.107,
+ 0.371,
+ 0.153,
+ 0.382
+ ],
+ "angle": 0,
+ "content": "(a) Input"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.203,
+ 0.273,
+ 0.304,
+ 0.352
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.23,
+ 0.354,
+ 0.278,
+ 0.364
+ ],
+ "angle": 0,
+ "content": "Inst:18"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.209,
+ 0.371,
+ 0.298,
+ 0.381
+ ],
+ "angle": 0,
+ "content": "(b) Ground Truth"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.328,
+ 0.273,
+ 0.429,
+ 0.351
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.358,
+ 0.354,
+ 0.398,
+ 0.364
+ ],
+ "angle": 0,
+ "content": "Inst:4"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.344,
+ 0.371,
+ 0.412,
+ 0.381
+ ],
+ "angle": 0,
+ "content": "(c) GeoTrans"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.452,
+ 0.273,
+ 0.553,
+ 0.351
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.482,
+ 0.354,
+ 0.522,
+ 0.364
+ ],
+ "angle": 0,
+ "content": "Inst:8"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.48,
+ 0.371,
+ 0.525,
+ 0.381
+ ],
+ "angle": 0,
+ "content": "(d) Ours"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.565,
+ 0.273,
+ 0.716,
+ 0.351
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.611,
+ 0.354,
+ 0.668,
+ 0.364
+ ],
+ "angle": 0,
+ "content": "IR:48.03%"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.59,
+ 0.371,
+ 0.69,
+ 0.381
+ ],
+ "angle": 0,
+ "content": "(e) GeoTrans (corr)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.732,
+ 0.273,
+ 0.884,
+ 0.351
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.78,
+ 0.354,
+ 0.838,
+ 0.364
+ ],
+ "angle": 0,
+ "content": "IR:72.40%"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.771,
+ 0.371,
+ 0.845,
+ 0.381
+ ],
+ "angle": 0,
+ "content": "(f) Ours (corr)"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.076,
+ 0.386,
+ 0.891,
+ 0.428
+ ],
+ "angle": 0,
+ "content": "Figure 5. Registration results on ROBI benchmark. We visualize the successfully registered instances in (c) and (d). MIRETR registers more instances in the cluttered scenes (the \\(2^{\\text{nd}}\\) and the \\(3^{\\text{rd}}\\) rows) and with incomplete geometry (all three rows). And it extracts more accurate correspondences benefiting from the instance-aware correspondence learning mechanism."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.08,
+ 0.434,
+ 0.466,
+ 0.603
+ ],
+ "angle": 0,
+ "content": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | 10.35 | 1.34 | 0.72 | 0.93 |
| CoFiNet [54] + RansaCov [34] | 1.25 | 1.39 | 1.31 |
| CoFiNet [54] + PointCLM [55] | 1.17 | 2.11 | 1.50 |
| CoFiNet [54] + ECC [44] | 3.21 | 8.50 | 4.66 |
| GeoTransformer [38] + T-Linkage [33] | 39.64 | 7.71 | 4.57 | 5.73 |
| GeoTransformer [38] + RansaCov [34] | 8.99 | 13.58 | 10.81 |
| GeoTransformer [38] + PointCLM [55] | 13.98 | 26.64 | 18.33 |
| GeoTransformer [38] + ECC [44] | 18.55 | 30.99 | 23.20 |
| MIRETR (ours) + T-Linkage [33] | 45.14 | 12.04 | 10.47 | 11.20 |
| MIRETR (ours) + RansaCov [34] | 14.14 | 26.29 | 18.38 |
| MIRETR (ours) + PointCLM [55] | 18.68 | 40.11 | 25.48 |
| MIRETR (ours) + ECC [44] | 24.65 | 34.85 | 28.91 |
| MIRETR (ours, full pipeline) | 38.51 | 41.19 | 39.80 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.128,
+ 0.608,
+ 0.417,
+ 0.62
+ ],
+ "angle": 0,
+ "content": "Table 2. Evaluation results on ROBI benchmark."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.628,
+ 0.468,
+ 0.733
+ ],
+ "angle": 0,
+ "content": "forms the baseline correspondence methods by a large margin on all four metrics. Compared to Scan2CAD, ROBI is more challenging due to the more cluttered scenes, where instance-aware information is more important. Benefiting from the instance-aware geometric transformer module, our method extracts reliable instance-aware correspondences, which contributes to more accurate instance registrations."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.735,
+ 0.468,
+ 0.87
+ ],
+ "angle": 0,
+ "content": "Our full model achieves significant improvements over the ECC-based model on MR, MP and MF. The PointCLM-based model achieves comparable MP with the full model, but significantly worse MR, which means it can register only several easy instances but misses most heavily-occluded instances. Specifically, it can register on average only 6.2 instances per scene, while the number of our method is 13.7. Our method achieves both high MR and MP, showing strong robustness in cluttered scenes."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.871,
+ 0.468,
+ 0.899
+ ],
+ "angle": 0,
+ "content": "Qualitative results. Fig. 5 visualizes the registration results of GeoTransformer and MIRETR. Our method can"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.503,
+ 0.433,
+ 0.627,
+ 0.53
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.631,
+ 0.433,
+ 0.757,
+ 0.53
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.763,
+ 0.433,
+ 0.889,
+ 0.53
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.502,
+ 0.532,
+ 0.629,
+ 0.629
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.507,
+ 0.63,
+ 0.625,
+ 0.64
+ ],
+ "angle": 0,
+ "content": "(a) Anchor's neighbors"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.632,
+ 0.532,
+ 0.758,
+ 0.629
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.637,
+ 0.63,
+ 0.755,
+ 0.64
+ ],
+ "angle": 0,
+ "content": "(b) Ground Truth Mask"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.762,
+ 0.532,
+ 0.889,
+ 0.629
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.764,
+ 0.63,
+ 0.89,
+ 0.64
+ ],
+ "angle": 0,
+ "content": "(c) Our Prediceted Mask"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.5,
+ 0.642,
+ 0.891,
+ 0.698
+ ],
+ "angle": 0,
+ "content": "Figure 6. Visualizations of the predicted instance masks. The anchor superpoint (patch) is in red and the selected neighbors are color-coded. Our method can effectively reject the patches outside the interested instance while preserving most of those inside."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.5,
+ 0.703,
+ 0.891,
+ 0.763
+ ],
+ "angle": 0,
+ "content": "register more instances in cluttered scenes (see the \\(2^{\\text{nd}}\\) and the \\(3^{\\text{rd}}\\) rows) and with severe geometric deficiency (see all three rows). And MIRETR attains more high-quality correspondences thanks to our instance-aware design."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.5,
+ 0.765,
+ 0.891,
+ 0.884
+ ],
+ "angle": 0,
+ "content": "We further visualize the predicted instance masks Fig. 6. The instance masks effectively reject the patches out of the interested instance while preserving most of those inside. Specifically, our method achieves the instance mIoU of 69.26 pp while the result of the model without masking is merely 38.13 pp. The high-quality masks allow us to effectively encode intra-instance context and contribute to accurate registrations."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.519,
+ 0.886,
+ 0.89,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Fig. 7 compares the results under different overlap ratios."
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.518,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "19611"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.078,
+ 0.087,
+ 0.473,
+ 0.265
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.078,
+ 0.266,
+ 0.468,
+ 0.28
+ ],
+ "angle": 0,
+ "content": "Figure 7. Results in different overlap levels on ROBI benchmark."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.288,
+ 0.47,
+ 0.35
+ ],
+ "angle": 0,
+ "content": "MIRETR consistently outperforms the baselines at all overlap levels, and the improvements are more significant when the overlap ratio is below \\(50\\%\\). This proves the strong capability of our method to handle low-overlap instances."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.359,
+ 0.308,
+ 0.376
+ ],
+ "angle": 0,
+ "content": "5.3. Evaluations on ShapeNet"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.383,
+ 0.47,
+ 0.534
+ ],
+ "angle": 0,
+ "content": "Dataset. ShapeNet [10] is a large CAD model dataset which contains 51300 models from 55 categories. We use the models from first 30 categories for training and those from the rest 25 categories for testing to evaluate the generality to novel categories. We randomly sample at most 500 models from each category to avoid class imbalance. For each point cloud pair, the model point cloud is a random CAD model and the scene one is constructed by applying \\(4\\sim 16\\) random poses on the model. At last, we obtain 8634 pairs for training, 900 for validation, and 7683 for testing."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.535,
+ 0.471,
+ 0.746
+ ],
+ "angle": 0,
+ "content": "Quantitative results. As shown in Tab. 3, our MIRETR achieves the best results on all four metrics, surpassing the baselines by a large margin. GeoTransformer significantly outperforms CoFiNet, which proves the importance of geometric structure to achieving generality. And our method achieves improvements of more than 10 pp over GeoTransformer on all registration metrics with PointCLM and ECC. Thanks to the instance-ware geometric transformer, the correspondences extracted by our method provide strong instance information, facilitating the correspondence clustering process. At last, our full model surpasses the previous best ECC-based baseline by around 2 pp on MR, 11 pp on MP and 7 pp on MF, demonstrating the strong generality of MIRETR to unseen objects."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.757,
+ 0.241,
+ 0.771
+ ],
+ "angle": 0,
+ "content": "5.4. Ablation Studies"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.78,
+ 0.47,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "We conduct extensive ablation studies on ROBI benchmark to better understand our design choices in Tab. 4. We first ablate the instance-aware point matching and extract point correspondences within only the patches of the superpoint correspondences in Tab. 4 (b), which is significantly worse than the full model. The poses estimated by this method is unstable due to clustered correspondences, thus the duplicated candidates cannot be effectively removed by NMS."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.503,
+ 0.089,
+ 0.89,
+ 0.259
+ ],
+ "angle": 0,
+ "content": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | | 4.59 | 3.24 | 3.79 |
| CoFiNet [54] + RansaCov [34] | 39.53 | 7.35 | 5.33 | 6.17 |
| CoFiNet [54] + PointCLM [55] | | 23.40 | 19.66 | 21.36 |
| CoFiNet [54] + ECC [44] | | 41.49 | 26.60 | 32.41 |
| GeoTransformer [38] + T-Linkage [33] | | 27.68 | 26.40 | 27.02 |
| GeoTransformer [38] + RansaCov [34] | 69.65 | 41.86 | 48.16 | 44.78 |
| GeoTransformer [38] + PointCLM [55] | | 68.09 | 69.09 | 68.58 |
| GeoTransformer [38] + ECC [44] | | 78.03 | 64.08 | 70.37 |
| MIRETR (ours) + T-Linkage [33] | | 29.40 | 28.67 | 29.03 |
| MIRETR (ours) + RansaCov [34] | | 45.63 | 46.85 | 46.23 |
| MIRETR (ours) + PointCLM [55] | 77.13 | 85.63 | 85.81 | 85.71 |
| MIRETR (ours) + ECC [44] | | 92.56 | 82.08 | 87.01 |
| MIRETR (ours, full pipeline) | | 94.95 | 93.94 | 94.44 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.54,
+ 0.264,
+ 0.851,
+ 0.277
+ ],
+ "angle": 0,
+ "content": "Table 3. Evaluation results on ShapeNet benchmark."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.505,
+ 0.28,
+ 0.887,
+ 0.356
+ ],
+ "angle": 0,
+ "content": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| FCGF [14] + T-Linkage [33] | | 54.75 | 22.76 | 32.15 |
| FCGF [14] + RansaCov [34] | 36.27 | 73.50 | 45.01 | 55.83 |
| FCGF [14] + PointCLM [55] | | 86.99 | 70.05 | 77.60 |
| FCGF [14] + ECC [44] | | 92.60 | 73.79 | 82.13 |
| CoFiNet [54] + T-Linkage [33] | | 28.17 | 5.09 | 8.62 |
| CoFiNet [54] + RansaCov [34] | 18.52 | 38.29 | 13.69 | 20.16 |
| CoFiNet [54] + PointCLM [55] | | 36.15 | 24.70 | 29.34 |
| CoFiNet [54] + ECC [44] | | 57.26 | 17.24 | 26.50 |
| GeoTransformer [38] + T-Linkage [33] | | 72.68 | 44.80 | 55.43 |
| GeoTransformer [38] + RansaCov [34] | 44.02 | 78.84 | 67.16 | 72.53 |
| GeoTransformer [38] + PointCLM [55] | | 82.81 | 81.90 | 82.35 |
| GeoTransformer [38] + ECC [44] | | 94.63 | 74.83 | 83.57 |
| MIRETR (ours) + T-Linkage [33] | | 77.12 | 46.04 | 57.65 |
| MIRETR (ours) + RansaCov [34] | | 84.78 | 71.34 | 77.48 |
| MIRETR (ours) + PointCLM [55] | 56.59 | 91.85 | 91.08 | 91.46 |
| MIRETR (ours) + ECC [44] | | 96.52 | 89.03 | 92.62 |
| MIRETR (ours, full pipeline) | | 95.70 | 91.21 | 93.40 |
+
+Table 1. Evaluation results on Scan2CAD benchmark.
+
+# 5.1. Evaluations on Scan2CAD
+
+Dataset. Scan2CAD [2] is a scene-to-CAD alignment dataset build upon ScanNet [15] and ShapeNet [10]. It consists of 1506 scenes from ScanNet annotated with 14225 CAD models from ShapeNet and their poses in the scenes. Following [55], we replace the objects in the scenes with the corresponding aligned CAD models. We select the scene-model pairs which the scene contain multiple instances of the model for multi-instance registration. As last, we obtain 2184 point cloud pairs and use $70\%$ pairs for training, $10\%$ for validation and $20\%$ for testing.
+
+Quantitative results. As shown in Tab. 1, our MIRETR achieves consistent improvements over the baseline correspondence models. Notably, our method surpasses the previous best GeoTransformer on IR by 12 percentage points (pp). When coupled with PointCLM, it outperforms GeoTransformer by over 9 pp on the three registration metrics.
+
+Moreover, our full pipeline surpasses all models with the baseline multi-model fitting methods. Although the ECC-based model performs slightly better than full MIRETR on MR, its MP is significantly worse. This indicates that ECC achieves high MR by predicting more registrations, which is impractical in real-world applications. And our method outperforms PointCLM on all three registration metrics, which demonstrates the strong effectiveness of our method.
+
+# 5.2. Evaluations on ROBI
+
+Dataset. ROBI [52] is a recent dataset for industrial bin-picking. It includes 7 reflective metallic industrial objects and 63 bin-picking scenes. Each scene contains cluttered instances of one industrial object. For each point cloud pair, the scene point cloud is backprojected from a depth image and the model point cloud is sampled from the CAD model of its corresponding industrial object. We obtain 4880 pairs in total, and we split the scenes for training, validation and testing according to the ratio of $6:1:2$ for each object.
+
+Quantitative results. As in Tab. 2, our method outper
+
+
+
+
+Inst:12
+
+
+
+
+
+
+
+
+
+
+
+
+Inst:22
+
+
+Inst:3
+
+
+Inst:9
+
+
+IR:39.42%
+IR:49.91%
+
+
+IR:54.21%
+IR:56.90%
+
+
+(a) Input
+
+
+Inst:18
+(b) Ground Truth
+
+
+Inst:4
+Inst:4
+(c) GeoTrans
+
+
+Inst:11
+Inst:8
+(d) Ours
+
+
+IR:48.03%
+(e) GeoTrans (corr)
+
+
+IR:72.40%
+(f) Ours (corr)
+
+| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | 10.35 | 1.34 | 0.72 | 0.93 |
| CoFiNet [54] + RansaCov [34] | 1.25 | 1.39 | 1.31 |
| CoFiNet [54] + PointCLM [55] | 1.17 | 2.11 | 1.50 |
| CoFiNet [54] + ECC [44] | 3.21 | 8.50 | 4.66 |
| GeoTransformer [38] + T-Linkage [33] | 39.64 | 7.71 | 4.57 | 5.73 |
| GeoTransformer [38] + RansaCov [34] | 8.99 | 13.58 | 10.81 |
| GeoTransformer [38] + PointCLM [55] | 13.98 | 26.64 | 18.33 |
| GeoTransformer [38] + ECC [44] | 18.55 | 30.99 | 23.20 |
| MIRETR (ours) + T-Linkage [33] | 45.14 | 12.04 | 10.47 | 11.20 |
| MIRETR (ours) + RansaCov [34] | 14.14 | 26.29 | 18.38 |
| MIRETR (ours) + PointCLM [55] | 18.68 | 40.11 | 25.48 |
| MIRETR (ours) + ECC [44] | 24.65 | 34.85 | 28.91 |
| MIRETR (ours, full pipeline) | 38.51 | 41.19 | 39.80 |
+
+Table 2. Evaluation results on ROBI benchmark.
+
+forms the baseline correspondence methods by a large margin on all four metrics. Compared to Scan2CAD, ROBI is more challenging due to the more cluttered scenes, where instance-aware information is more important. Benefiting from the instance-aware geometric transformer module, our method extracts reliable instance-aware correspondences, which contributes to more accurate instance registrations.
+
+Our full model achieves significant improvements over the ECC-based model on MR, MP and MF. The PointCLM-based model achieves comparable MP with the full model, but significantly worse MR, which means it can register only several easy instances but misses most heavily-occluded instances. Specifically, it can register on average only 6.2 instances per scene, while the number of our method is 13.7. Our method achieves both high MR and MP, showing strong robustness in cluttered scenes.
+
+Qualitative results. Fig. 5 visualizes the registration results of GeoTransformer and MIRETR. Our method can
+
+
+Figure 5. Registration results on ROBI benchmark. We visualize the successfully registered instances in (c) and (d). MIRETR registers more instances in the cluttered scenes (the $2^{\text{nd}}$ and the $3^{\text{rd}}$ rows) and with incomplete geometry (all three rows). And it extracts more accurate correspondences benefiting from the instance-aware correspondence learning mechanism.
+
+
+
+
+
+
+(a) Anchor's neighbors
+Figure 6. Visualizations of the predicted instance masks. The anchor superpoint (patch) is in red and the selected neighbors are color-coded. Our method can effectively reject the patches outside the interested instance while preserving most of those inside.
+
+
+(b) Ground Truth Mask
+
+
+(c) Our Prediceted Mask
+
+register more instances in cluttered scenes (see the $2^{\text{nd}}$ and the $3^{\text{rd}}$ rows) and with severe geometric deficiency (see all three rows). And MIRETR attains more high-quality correspondences thanks to our instance-aware design.
+
+We further visualize the predicted instance masks Fig. 6. The instance masks effectively reject the patches out of the interested instance while preserving most of those inside. Specifically, our method achieves the instance mIoU of 69.26 pp while the result of the model without masking is merely 38.13 pp. The high-quality masks allow us to effectively encode intra-instance context and contribute to accurate registrations.
+
+Fig. 7 compares the results under different overlap ratios.
+
+
+Figure 7. Results in different overlap levels on ROBI benchmark.
+
+MIRETR consistently outperforms the baselines at all overlap levels, and the improvements are more significant when the overlap ratio is below $50\%$ . This proves the strong capability of our method to handle low-overlap instances.
+
+# 5.3. Evaluations on ShapeNet
+
+Dataset. ShapeNet [10] is a large CAD model dataset which contains 51300 models from 55 categories. We use the models from first 30 categories for training and those from the rest 25 categories for testing to evaluate the generality to novel categories. We randomly sample at most 500 models from each category to avoid class imbalance. For each point cloud pair, the model point cloud is a random CAD model and the scene one is constructed by applying $4\sim 16$ random poses on the model. At last, we obtain 8634 pairs for training, 900 for validation, and 7683 for testing.
+
+Quantitative results. As shown in Tab. 3, our MIRETR achieves the best results on all four metrics, surpassing the baselines by a large margin. GeoTransformer significantly outperforms CoFiNet, which proves the importance of geometric structure to achieving generality. And our method achieves improvements of more than 10 pp over GeoTransformer on all registration metrics with PointCLM and ECC. Thanks to the instance-ware geometric transformer, the correspondences extracted by our method provide strong instance information, facilitating the correspondence clustering process. At last, our full model surpasses the previous best ECC-based baseline by around 2 pp on MR, 11 pp on MP and 7 pp on MF, demonstrating the strong generality of MIRETR to unseen objects.
+
+# 5.4. Ablation Studies
+
+We conduct extensive ablation studies on ROBI benchmark to better understand our design choices in Tab. 4. We first ablate the instance-aware point matching and extract point correspondences within only the patches of the superpoint correspondences in Tab. 4 (b), which is significantly worse than the full model. The poses estimated by this method is unstable due to clustered correspondences, thus the duplicated candidates cannot be effectively removed by NMS.
+
+| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | | 4.59 | 3.24 | 3.79 |
| CoFiNet [54] + RansaCov [34] | 39.53 | 7.35 | 5.33 | 6.17 |
| CoFiNet [54] + PointCLM [55] | | 23.40 | 19.66 | 21.36 |
| CoFiNet [54] + ECC [44] | | 41.49 | 26.60 | 32.41 |
| GeoTransformer [38] + T-Linkage [33] | | 27.68 | 26.40 | 27.02 |
| GeoTransformer [38] + RansaCov [34] | 69.65 | 41.86 | 48.16 | 44.78 |
| GeoTransformer [38] + PointCLM [55] | | 68.09 | 69.09 | 68.58 |
| GeoTransformer [38] + ECC [44] | | 78.03 | 64.08 | 70.37 |
| MIRETR (ours) + T-Linkage [33] | | 29.40 | 28.67 | 29.03 |
| MIRETR (ours) + RansaCov [34] | | 45.63 | 46.85 | 46.23 |
| MIRETR (ours) + PointCLM [55] | 77.13 | 85.63 | 85.81 | 85.71 |
| MIRETR (ours) + ECC [44] | | 92.56 | 82.08 | 87.01 |
| MIRETR (ours, full pipeline) | | 94.95 | 93.94 | 94.44 |
+
+Table 3. Evaluation results on ShapeNet benchmark.
+
+| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| FCGF [14] + T-Linkage [33] | | 54.75 | 22.76 | 32.15 |
| FCGF [14] + RansaCov [34] | 36.27 | 73.50 | 45.01 | 55.83 |
| FCGF [14] + PointCLM [55] | | 86.99 | 70.05 | 77.60 |
| FCGF [14] + ECC [44] | | 92.60 | 73.79 | 82.13 |
| CoFiNet [54] + T-Linkage [33] | | 28.17 | 5.09 | 8.62 |
| CoFiNet [54] + RansaCov [34] | 18.52 | 38.29 | 13.69 | 20.16 |
| CoFiNet [54] + PointCLM [55] | | 36.15 | 24.70 | 29.34 |
| CoFiNet [54] + ECC [44] | | 57.26 | 17.24 | 26.50 |
| GeoTransformer [38] + T-Linkage [33] | | 72.68 | 44.80 | 55.43 |
| GeoTransformer [38] + RansaCov [34] | 44.02 | 78.84 | 67.16 | 72.53 |
| GeoTransformer [38] + PointCLM [55] | | 82.81 | 81.90 | 82.35 |
| GeoTransformer [38] + ECC [44] | | 94.63 | 74.83 | 83.57 |
| MIRETR (ours) + T-Linkage [33] | | 77.12 | 46.04 | 57.65 |
| MIRETR (ours) + RansaCov [34] | | 84.78 | 71.34 | 77.48 |
| MIRETR (ours) + PointCLM [55] | 56.59 | 91.85 | 91.08 | 91.46 |
| MIRETR (ours) + ECC [44] | | 96.52 | 89.03 | 92.62 |
| MIRETR (ours, full pipeline) | | 95.70 | 91.21 | 93.40 |
",
+ "image_path": "6861a70b9f788145571310bf2c171c61f4a4da672434117e90d1d85f2b30f5c4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 326,
+ 245,
+ 523,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 326,
+ 245,
+ 523,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 326,
+ 245,
+ 523,
+ 255
+ ],
+ "type": "text",
+ "content": "Table 1. Evaluation results on Scan2CAD benchmark."
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 306,
+ 262,
+ 454,
+ 274
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 262,
+ 454,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 262,
+ 454,
+ 274
+ ],
+ "type": "text",
+ "content": "5.1. Evaluations on Scan2CAD"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "text",
+ "content": "Dataset. Scan2CAD [2] is a scene-to-CAD alignment dataset build upon ScanNet [15] and ShapeNet [10]. It consists of 1506 scenes from ScanNet annotated with 14225 CAD models from ShapeNet and their poses in the scenes. Following [55], we replace the objects in the scenes with the corresponding aligned CAD models. We select the scene-model pairs which the scene contain multiple instances of the model for multi-instance registration. As last, we obtain 2184 point cloud pairs and use "
+ },
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "inline_equation",
+ "content": "70\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "text",
+ "content": " pairs for training, "
+ },
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "inline_equation",
+ "content": "10\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "text",
+ "content": " for validation and "
+ },
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "inline_equation",
+ "content": "20\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 281,
+ 545,
+ 399
+ ],
+ "type": "text",
+ "content": " for testing."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 304,
+ 400,
+ 545,
+ 472
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 400,
+ 545,
+ 472
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 400,
+ 545,
+ 472
+ ],
+ "type": "text",
+ "content": "Quantitative results. As shown in Tab. 1, our MIRETR achieves consistent improvements over the baseline correspondence models. Notably, our method surpasses the previous best GeoTransformer on IR by 12 percentage points (pp). When coupled with PointCLM, it outperforms GeoTransformer by over 9 pp on the three registration metrics."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 304,
+ 472,
+ 545,
+ 567
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 472,
+ 545,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 472,
+ 545,
+ 567
+ ],
+ "type": "text",
+ "content": "Moreover, our full pipeline surpasses all models with the baseline multi-model fitting methods. Although the ECC-based model performs slightly better than full MIRETR on MR, its MP is significantly worse. This indicates that ECC achieves high MR by predicting more registrations, which is impractical in real-world applications. And our method outperforms PointCLM on all three registration metrics, which demonstrates the strong effectiveness of our method."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 306,
+ 575,
+ 429,
+ 586
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 575,
+ 429,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 575,
+ 429,
+ 586
+ ],
+ "type": "text",
+ "content": "5.2. Evaluations on ROBI"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 304,
+ 594,
+ 545,
+ 701
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 594,
+ 545,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 594,
+ 545,
+ 701
+ ],
+ "type": "text",
+ "content": "Dataset. ROBI [52] is a recent dataset for industrial bin-picking. It includes 7 reflective metallic industrial objects and 63 bin-picking scenes. Each scene contains cluttered instances of one industrial object. For each point cloud pair, the scene point cloud is backprojected from a depth image and the model point cloud is sampled from the CAD model of its corresponding industrial object. We obtain 4880 pairs in total, and we split the scenes for training, validation and testing according to the ratio of "
+ },
+ {
+ "bbox": [
+ 304,
+ 594,
+ 545,
+ 701
+ ],
+ "type": "inline_equation",
+ "content": "6:1:2"
+ },
+ {
+ "bbox": [
+ 304,
+ 594,
+ 545,
+ 701
+ ],
+ "type": "text",
+ "content": " for each object."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 306,
+ 702,
+ 545,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 702,
+ 545,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 702,
+ 545,
+ 714
+ ],
+ "type": "text",
+ "content": "Quantitative results. As in Tab. 2, our method outper"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "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": "19610"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 49,
+ 68,
+ 110,
+ 130
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 49,
+ 68,
+ 110,
+ 130
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 68,
+ 110,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 68,
+ 110,
+ 130
+ ],
+ "type": "image",
+ "image_path": "1abb0532efa0b1f690a72759bce9d97f471e82cd995a9373c97e80cf534597da.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 125,
+ 68,
+ 186,
+ 130
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 125,
+ 68,
+ 186,
+ 130
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 125,
+ 68,
+ 186,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 125,
+ 68,
+ 186,
+ 130
+ ],
+ "type": "image",
+ "image_path": "704b66b88c6729d9e6d53705b971c818475de3b48203da143327ec604af1982b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 140,
+ 132,
+ 170,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 132,
+ 170,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 132,
+ 170,
+ 140
+ ],
+ "type": "text",
+ "content": "Inst:12"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 200,
+ 68,
+ 261,
+ 129
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 200,
+ 68,
+ 261,
+ 129
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 68,
+ 261,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 68,
+ 261,
+ 129
+ ],
+ "type": "image",
+ "image_path": "644149e1b7874c39a7f61a04f1fbf69b621909165d2e10ea50794c9885073d18.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 276,
+ 68,
+ 337,
+ 129
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 276,
+ 68,
+ 337,
+ 129
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 68,
+ 337,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 68,
+ 337,
+ 129
+ ],
+ "type": "image",
+ "image_path": "802cd1095a9446fc722eebc795ebfe0682251baab7a87e94de02204906d214f5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 345,
+ 68,
+ 440,
+ 129
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 345,
+ 68,
+ 440,
+ 129
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 345,
+ 68,
+ 440,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 68,
+ 440,
+ 129
+ ],
+ "type": "image",
+ "image_path": "98d47ff7e66ebf5c4a5024041eec2b8facef842f7d79a9b042ef23194aeed501.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 447,
+ 68,
+ 542,
+ 130
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 447,
+ 68,
+ 542,
+ 130
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 447,
+ 68,
+ 542,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 447,
+ 68,
+ 542,
+ 130
+ ],
+ "type": "image",
+ "image_path": "dcf490464ac34f05aef2449727051e01a1176447869aa82803f6a3b80d6ebae8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 49,
+ 141,
+ 110,
+ 202
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 49,
+ 141,
+ 110,
+ 202
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 49,
+ 141,
+ 110,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 49,
+ 141,
+ 110,
+ 202
+ ],
+ "type": "image",
+ "image_path": "69bf9dfd46a5d4c4b990e3f08d9f2ceea561adb39a0152fe649cafaaf2d4b944.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 124,
+ 141,
+ 186,
+ 202
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 124,
+ 141,
+ 186,
+ 202
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 124,
+ 141,
+ 186,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 141,
+ 186,
+ 202
+ ],
+ "type": "image",
+ "image_path": "a51edfd444ebd817b90e9e521858ae6e30c964e1c3b3c8b17636ec74df0ec497.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 140,
+ 205,
+ 170,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 205,
+ 170,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 205,
+ 170,
+ 213
+ ],
+ "type": "text",
+ "content": "Inst:22"
+ }
+ ]
+ }
+ ],
+ "index": 13,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 201,
+ 141,
+ 261,
+ 202
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 219,
+ 132,
+ 243,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 132,
+ 243,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 132,
+ 243,
+ 140
+ ],
+ "type": "text",
+ "content": "Inst:3"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 201,
+ 141,
+ 261,
+ 202
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 201,
+ 141,
+ 261,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 201,
+ 141,
+ 261,
+ 202
+ ],
+ "type": "image",
+ "image_path": "61f2e6e3fb21f46a27a0863d5a0a1fe1e21d5f959140175d540a28ac1c232176.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 14,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 276,
+ 141,
+ 337,
+ 202
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 294,
+ 132,
+ 319,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 132,
+ 319,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 132,
+ 319,
+ 140
+ ],
+ "type": "text",
+ "content": "Inst:9"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 276,
+ 141,
+ 337,
+ 202
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 141,
+ 337,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 141,
+ 337,
+ 202
+ ],
+ "type": "image",
+ "image_path": "47222ca7df42dc2273d530c4ccfaf40b595aeb40c9ba23287a6c505bda959a24.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 353,
+ 141,
+ 432,
+ 202
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 373,
+ 132,
+ 409,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 373,
+ 132,
+ 409,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 373,
+ 132,
+ 409,
+ 140
+ ],
+ "type": "text",
+ "content": "IR:39.42%"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 353,
+ 141,
+ 432,
+ 202
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 353,
+ 141,
+ 432,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 353,
+ 141,
+ 432,
+ 202
+ ],
+ "type": "image",
+ "image_path": "9a7fcfb515a5db074b12478da6599137d1ac11a0ae2da99a77456ceb4d27c2a0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 18,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 373,
+ 205,
+ 408,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 373,
+ 205,
+ 408,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 373,
+ 205,
+ 408,
+ 213
+ ],
+ "type": "text",
+ "content": "IR:49.91%"
+ }
+ ]
+ }
+ ],
+ "index": 19,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 455,
+ 141,
+ 534,
+ 202
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 477,
+ 132,
+ 512,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 132,
+ 512,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 132,
+ 512,
+ 140
+ ],
+ "type": "text",
+ "content": "IR:54.21%"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 455,
+ 141,
+ 534,
+ 202
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 455,
+ 141,
+ 534,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 455,
+ 141,
+ 534,
+ 202
+ ],
+ "type": "image",
+ "image_path": "251cbcaee9afec1b852c631e6866a3942684bba4cd67373ccb205ee2290cc1e5.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 20,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 477,
+ 205,
+ 512,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 205,
+ 512,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 205,
+ 512,
+ 213
+ ],
+ "type": "text",
+ "content": "IR:56.90%"
+ }
+ ]
+ }
+ ],
+ "index": 21,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 48,
+ 215,
+ 111,
+ 278
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 48,
+ 215,
+ 111,
+ 278
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 215,
+ 111,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 215,
+ 111,
+ 278
+ ],
+ "type": "image",
+ "image_path": "b69c08aeb868cc2fad1379dd2651a4718e7043dc9f9bed2dccc4cd7a2d810a69.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 22,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 65,
+ 293,
+ 93,
+ 302
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 65,
+ 293,
+ 93,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 65,
+ 293,
+ 93,
+ 302
+ ],
+ "type": "text",
+ "content": "(a) Input"
+ }
+ ]
+ }
+ ],
+ "index": 23,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 124,
+ 216,
+ 186,
+ 278
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 124,
+ 216,
+ 186,
+ 278
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 124,
+ 216,
+ 186,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 216,
+ 186,
+ 278
+ ],
+ "type": "image",
+ "image_path": "88ba29007d8ff52a2871dd70323d5b0a20822679cfaf53c00c1570e020f5dcf1.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 24,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 140,
+ 280,
+ 170,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 140,
+ 280,
+ 170,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 280,
+ 170,
+ 288
+ ],
+ "type": "text",
+ "content": "Inst:18"
+ }
+ ]
+ }
+ ],
+ "index": 25,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 127,
+ 293,
+ 182,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 127,
+ 293,
+ 182,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 127,
+ 293,
+ 182,
+ 301
+ ],
+ "type": "text",
+ "content": "(b) Ground Truth"
+ }
+ ]
+ }
+ ],
+ "index": 26,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 200,
+ 216,
+ 262,
+ 277
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 219,
+ 205,
+ 243,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 205,
+ 243,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 205,
+ 243,
+ 213
+ ],
+ "type": "text",
+ "content": "Inst:4"
+ }
+ ]
+ }
+ ],
+ "index": 15,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 200,
+ 216,
+ 262,
+ 277
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 216,
+ 262,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 216,
+ 262,
+ 277
+ ],
+ "type": "image",
+ "image_path": "d018345a684d4f108de169b40602c60da8d37b88e7b148d84d70253f3f89d44f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 27,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 219,
+ 280,
+ 243,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 219,
+ 280,
+ 243,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 219,
+ 280,
+ 243,
+ 288
+ ],
+ "type": "text",
+ "content": "Inst:4"
+ }
+ ]
+ }
+ ],
+ "index": 28,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 210,
+ 293,
+ 252,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 210,
+ 293,
+ 252,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 210,
+ 293,
+ 252,
+ 301
+ ],
+ "type": "text",
+ "content": "(c) GeoTrans"
+ }
+ ]
+ }
+ ],
+ "index": 29,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 276,
+ 216,
+ 338,
+ 277
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 293,
+ 205,
+ 320,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 205,
+ 320,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 205,
+ 320,
+ 213
+ ],
+ "type": "text",
+ "content": "Inst:11"
+ }
+ ]
+ }
+ ],
+ "index": 17,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 276,
+ 216,
+ 338,
+ 277
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 216,
+ 338,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 216,
+ 338,
+ 277
+ ],
+ "type": "image",
+ "image_path": "4e921912da9654eb7fc66703f378cf70ad42ea341c3022fb1f929244e3610718.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 30,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 294,
+ 280,
+ 319,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 294,
+ 280,
+ 319,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 294,
+ 280,
+ 319,
+ 288
+ ],
+ "type": "text",
+ "content": "Inst:8"
+ }
+ ]
+ }
+ ],
+ "index": 31,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 293,
+ 293,
+ 321,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 293,
+ 321,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 293,
+ 321,
+ 301
+ ],
+ "type": "text",
+ "content": "(d) Ours"
+ }
+ ]
+ }
+ ],
+ "index": 32,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 345,
+ 216,
+ 438,
+ 277
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 345,
+ 216,
+ 438,
+ 277
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 345,
+ 216,
+ 438,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 345,
+ 216,
+ 438,
+ 277
+ ],
+ "type": "image",
+ "image_path": "1c725f11c388f00e1a1af134e30d50ede5a2a17ff74c64aca2d5f3b1480fa8b4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 33,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 373,
+ 280,
+ 408,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 373,
+ 280,
+ 408,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 373,
+ 280,
+ 408,
+ 288
+ ],
+ "type": "text",
+ "content": "IR:48.03%"
+ }
+ ]
+ }
+ ],
+ "index": 34,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 361,
+ 293,
+ 422,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 361,
+ 293,
+ 422,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 361,
+ 293,
+ 422,
+ 301
+ ],
+ "type": "text",
+ "content": "(e) GeoTrans (corr)"
+ }
+ ]
+ }
+ ],
+ "index": 35,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 447,
+ 216,
+ 541,
+ 277
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 447,
+ 216,
+ 541,
+ 277
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 447,
+ 216,
+ 541,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 447,
+ 216,
+ 541,
+ 277
+ ],
+ "type": "image",
+ "image_path": "289dc0081fe5c9ddbfe53c9c7fbf6005030cc6947efdd95e0d5d9d14a572b8ed.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 36,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 477,
+ 280,
+ 512,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 477,
+ 280,
+ 512,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 477,
+ 280,
+ 512,
+ 288
+ ],
+ "type": "text",
+ "content": "IR:72.40%"
+ }
+ ]
+ }
+ ],
+ "index": 37,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 471,
+ 293,
+ 517,
+ 301
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 471,
+ 293,
+ 517,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 471,
+ 293,
+ 517,
+ 301
+ ],
+ "type": "text",
+ "content": "(f) Ours (corr)"
+ }
+ ]
+ }
+ ],
+ "index": 38,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 48,
+ 343,
+ 285,
+ 477
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 48,
+ 343,
+ 285,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 48,
+ 343,
+ 285,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 48,
+ 343,
+ 285,
+ 477
+ ],
+ "type": "table",
+ "html": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | 10.35 | 1.34 | 0.72 | 0.93 |
| CoFiNet [54] + RansaCov [34] | 1.25 | 1.39 | 1.31 |
| CoFiNet [54] + PointCLM [55] | 1.17 | 2.11 | 1.50 |
| CoFiNet [54] + ECC [44] | 3.21 | 8.50 | 4.66 |
| GeoTransformer [38] + T-Linkage [33] | 39.64 | 7.71 | 4.57 | 5.73 |
| GeoTransformer [38] + RansaCov [34] | 8.99 | 13.58 | 10.81 |
| GeoTransformer [38] + PointCLM [55] | 13.98 | 26.64 | 18.33 |
| GeoTransformer [38] + ECC [44] | 18.55 | 30.99 | 23.20 |
| MIRETR (ours) + T-Linkage [33] | 45.14 | 12.04 | 10.47 | 11.20 |
| MIRETR (ours) + RansaCov [34] | 14.14 | 26.29 | 18.38 |
| MIRETR (ours) + PointCLM [55] | 18.68 | 40.11 | 25.48 |
| MIRETR (ours) + ECC [44] | 24.65 | 34.85 | 28.91 |
| MIRETR (ours, full pipeline) | 38.51 | 41.19 | 39.80 |
",
+ "image_path": "b12e836e86784d456a97df0e8d248d366226a281417fc839953ffa70b7175de9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 40,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 78,
+ 481,
+ 255,
+ 491
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 78,
+ 481,
+ 255,
+ 491
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 78,
+ 481,
+ 255,
+ 491
+ ],
+ "type": "text",
+ "content": "Table 2. Evaluation results on ROBI benchmark."
+ }
+ ]
+ }
+ ],
+ "index": 41,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 46,
+ 497,
+ 286,
+ 580
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 497,
+ 286,
+ 580
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 497,
+ 286,
+ 580
+ ],
+ "type": "text",
+ "content": "forms the baseline correspondence methods by a large margin on all four metrics. Compared to Scan2CAD, ROBI is more challenging due to the more cluttered scenes, where instance-aware information is more important. Benefiting from the instance-aware geometric transformer module, our method extracts reliable instance-aware correspondences, which contributes to more accurate instance registrations."
+ }
+ ]
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 46,
+ 582,
+ 286,
+ 689
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 582,
+ 286,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 582,
+ 286,
+ 689
+ ],
+ "type": "text",
+ "content": "Our full model achieves significant improvements over the ECC-based model on MR, MP and MF. The PointCLM-based model achieves comparable MP with the full model, but significantly worse MR, which means it can register only several easy instances but misses most heavily-occluded instances. Specifically, it can register on average only 6.2 instances per scene, while the number of our method is 13.7. Our method achieves both high MR and MP, showing strong robustness in cluttered scenes."
+ }
+ ]
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 46,
+ 689,
+ 286,
+ 712
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 689,
+ 286,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 689,
+ 286,
+ 712
+ ],
+ "type": "text",
+ "content": "Qualitative results. Fig. 5 visualizes the registration results of GeoTransformer and MIRETR. Our method can"
+ }
+ ]
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 307,
+ 342,
+ 383,
+ 419
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 46,
+ 305,
+ 545,
+ 338
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 305,
+ 545,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 305,
+ 545,
+ 338
+ ],
+ "type": "text",
+ "content": "Figure 5. Registration results on ROBI benchmark. We visualize the successfully registered instances in (c) and (d). MIRETR registers more instances in the cluttered scenes (the "
+ },
+ {
+ "bbox": [
+ 46,
+ 305,
+ 545,
+ 338
+ ],
+ "type": "inline_equation",
+ "content": "2^{\\text{nd}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 305,
+ 545,
+ 338
+ ],
+ "type": "text",
+ "content": " and the "
+ },
+ {
+ "bbox": [
+ 46,
+ 305,
+ 545,
+ 338
+ ],
+ "type": "inline_equation",
+ "content": "3^{\\text{rd}}"
+ },
+ {
+ "bbox": [
+ 46,
+ 305,
+ 545,
+ 338
+ ],
+ "type": "text",
+ "content": " rows) and with incomplete geometry (all three rows). And it extracts more accurate correspondences benefiting from the instance-aware correspondence learning mechanism."
+ }
+ ]
+ }
+ ],
+ "index": 39,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 307,
+ 342,
+ 383,
+ 419
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 342,
+ 383,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 342,
+ 383,
+ 419
+ ],
+ "type": "image",
+ "image_path": "391ca0191f955cb9a6ce937cf227c2f26566bf36dc2ed686e4e35c93a389f6c0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 45,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 386,
+ 342,
+ 463,
+ 419
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 386,
+ 342,
+ 463,
+ 419
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 386,
+ 342,
+ 463,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 342,
+ 463,
+ 419
+ ],
+ "type": "image",
+ "image_path": "b6485a00d3b1d395f9bd361f2487dcd38fa3167e21cd81dbb7ffe8c613cfcf55.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 46,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 466,
+ 342,
+ 544,
+ 419
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 466,
+ 342,
+ 544,
+ 419
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 466,
+ 342,
+ 544,
+ 419
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 466,
+ 342,
+ 544,
+ 419
+ ],
+ "type": "image",
+ "image_path": "d5503b738ba1ee730cc57be49e2512510deae3912135aa643abd992b3e952ba8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 47,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 307,
+ 421,
+ 384,
+ 498
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 421,
+ 384,
+ 498
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 421,
+ 384,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 421,
+ 384,
+ 498
+ ],
+ "type": "image",
+ "image_path": "8d3df5aa3f8921044baebe2827a67c205e342787a90d0bc6f09a89fe3001a9b0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 48,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 310,
+ 498,
+ 382,
+ 506
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 310,
+ 498,
+ 382,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 498,
+ 382,
+ 506
+ ],
+ "type": "text",
+ "content": "(a) Anchor's neighbors"
+ }
+ ]
+ }
+ ],
+ "index": 49,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 306,
+ 508,
+ 545,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 508,
+ 545,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 508,
+ 545,
+ 552
+ ],
+ "type": "text",
+ "content": "Figure 6. Visualizations of the predicted instance masks. The anchor superpoint (patch) is in red and the selected neighbors are color-coded. Our method can effectively reject the patches outside the interested instance while preserving most of those inside."
+ }
+ ]
+ }
+ ],
+ "index": 54,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 386,
+ 421,
+ 463,
+ 498
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 386,
+ 421,
+ 463,
+ 498
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 386,
+ 421,
+ 463,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 386,
+ 421,
+ 463,
+ 498
+ ],
+ "type": "image",
+ "image_path": "d5e8662215981f9b8f3c9bdfbf1940a9049c4a266abfef58d8dc728417113bbf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 50,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 389,
+ 498,
+ 462,
+ 506
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 389,
+ 498,
+ 462,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 389,
+ 498,
+ 462,
+ 506
+ ],
+ "type": "text",
+ "content": "(b) Ground Truth Mask"
+ }
+ ]
+ }
+ ],
+ "index": 51,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 466,
+ 421,
+ 544,
+ 498
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 466,
+ 421,
+ 544,
+ 498
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 466,
+ 421,
+ 544,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 466,
+ 421,
+ 544,
+ 498
+ ],
+ "type": "image",
+ "image_path": "96232d5edff7c4e3fb917f7bd8549b8ca128637140de62e2cc227b322af338d4.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 52,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 467,
+ 498,
+ 544,
+ 506
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 467,
+ 498,
+ 544,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 467,
+ 498,
+ 544,
+ 506
+ ],
+ "type": "text",
+ "content": "(c) Our Prediceted Mask"
+ }
+ ]
+ }
+ ],
+ "index": 53,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 306,
+ 556,
+ 545,
+ 604
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 556,
+ 545,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 556,
+ 545,
+ 604
+ ],
+ "type": "text",
+ "content": "register more instances in cluttered scenes (see the "
+ },
+ {
+ "bbox": [
+ 306,
+ 556,
+ 545,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "2^{\\text{nd}}"
+ },
+ {
+ "bbox": [
+ 306,
+ 556,
+ 545,
+ 604
+ ],
+ "type": "text",
+ "content": " and the "
+ },
+ {
+ "bbox": [
+ 306,
+ 556,
+ 545,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "3^{\\text{rd}}"
+ },
+ {
+ "bbox": [
+ 306,
+ 556,
+ 545,
+ 604
+ ],
+ "type": "text",
+ "content": " rows) and with severe geometric deficiency (see all three rows). And MIRETR attains more high-quality correspondences thanks to our instance-aware design."
+ }
+ ]
+ }
+ ],
+ "index": 55
+ },
+ {
+ "bbox": [
+ 306,
+ 605,
+ 545,
+ 700
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 605,
+ 545,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 605,
+ 545,
+ 700
+ ],
+ "type": "text",
+ "content": "We further visualize the predicted instance masks Fig. 6. The instance masks effectively reject the patches out of the interested instance while preserving most of those inside. Specifically, our method achieves the instance mIoU of 69.26 pp while the result of the model without masking is merely 38.13 pp. The high-quality masks allow us to effectively encode intra-instance context and contribute to accurate registrations."
+ }
+ ]
+ }
+ ],
+ "index": 56
+ },
+ {
+ "bbox": [
+ 317,
+ 701,
+ 544,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 317,
+ 701,
+ 544,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 317,
+ 701,
+ 544,
+ 713
+ ],
+ "type": "text",
+ "content": "Fig. 7 compares the results under different overlap ratios."
+ }
+ ]
+ }
+ ],
+ "index": 57
+ }
+ ],
+ "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": "19611"
+ }
+ ]
+ }
+ ],
+ "index": 58
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 47,
+ 68,
+ 289,
+ 209
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 47,
+ 68,
+ 289,
+ 209
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 68,
+ 289,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 68,
+ 289,
+ 209
+ ],
+ "type": "image",
+ "image_path": "47fd1ca19d7110631edeade7b18fcf7f75922bb74cb3687da7de3747fb0f468c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 47,
+ 210,
+ 286,
+ 221
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 210,
+ 286,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 210,
+ 286,
+ 221
+ ],
+ "type": "text",
+ "content": "Figure 7. Results in different overlap levels on ROBI benchmark."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 47,
+ 228,
+ 287,
+ 277
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 228,
+ 287,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 228,
+ 287,
+ 277
+ ],
+ "type": "text",
+ "content": "MIRETR consistently outperforms the baselines at all overlap levels, and the improvements are more significant when the overlap ratio is below "
+ },
+ {
+ "bbox": [
+ 47,
+ 228,
+ 287,
+ 277
+ ],
+ "type": "inline_equation",
+ "content": "50\\%"
+ },
+ {
+ "bbox": [
+ 47,
+ 228,
+ 287,
+ 277
+ ],
+ "type": "text",
+ "content": ". This proves the strong capability of our method to handle low-overlap instances."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 47,
+ 284,
+ 188,
+ 297
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 284,
+ 188,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 284,
+ 188,
+ 297
+ ],
+ "type": "text",
+ "content": "5.3. Evaluations on ShapeNet"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 287,
+ 422
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 303,
+ 287,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 303,
+ 287,
+ 422
+ ],
+ "type": "text",
+ "content": "Dataset. ShapeNet [10] is a large CAD model dataset which contains 51300 models from 55 categories. We use the models from first 30 categories for training and those from the rest 25 categories for testing to evaluate the generality to novel categories. We randomly sample at most 500 models from each category to avoid class imbalance. For each point cloud pair, the model point cloud is a random CAD model and the scene one is constructed by applying "
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 287,
+ 422
+ ],
+ "type": "inline_equation",
+ "content": "4\\sim 16"
+ },
+ {
+ "bbox": [
+ 46,
+ 303,
+ 287,
+ 422
+ ],
+ "type": "text",
+ "content": " random poses on the model. At last, we obtain 8634 pairs for training, 900 for validation, and 7683 for testing."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 46,
+ 423,
+ 288,
+ 590
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 423,
+ 288,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 423,
+ 288,
+ 590
+ ],
+ "type": "text",
+ "content": "Quantitative results. As shown in Tab. 3, our MIRETR achieves the best results on all four metrics, surpassing the baselines by a large margin. GeoTransformer significantly outperforms CoFiNet, which proves the importance of geometric structure to achieving generality. And our method achieves improvements of more than 10 pp over GeoTransformer on all registration metrics with PointCLM and ECC. Thanks to the instance-ware geometric transformer, the correspondences extracted by our method provide strong instance information, facilitating the correspondence clustering process. At last, our full model surpasses the previous best ECC-based baseline by around 2 pp on MR, 11 pp on MP and 7 pp on MF, demonstrating the strong generality of MIRETR to unseen objects."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 47,
+ 599,
+ 147,
+ 610
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 599,
+ 147,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 599,
+ 147,
+ 610
+ ],
+ "type": "text",
+ "content": "5.4. Ablation Studies"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 46,
+ 617,
+ 287,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 617,
+ 287,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 617,
+ 287,
+ 715
+ ],
+ "type": "text",
+ "content": "We conduct extensive ablation studies on ROBI benchmark to better understand our design choices in Tab. 4. We first ablate the instance-aware point matching and extract point correspondences within only the patches of the superpoint correspondences in Tab. 4 (b), which is significantly worse than the full model. The poses estimated by this method is unstable due to clustered correspondences, thus the duplicated candidates cannot be effectively removed by NMS."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 307,
+ 70,
+ 544,
+ 205
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 70,
+ 544,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 70,
+ 544,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 70,
+ 544,
+ 205
+ ],
+ "type": "table",
+ "html": "| Model | IR(%) | MR(%) | MP(%) | MF(%) |
| CoFiNet [54] + T-Linkage [33] | | 4.59 | 3.24 | 3.79 |
| CoFiNet [54] + RansaCov [34] | 39.53 | 7.35 | 5.33 | 6.17 |
| CoFiNet [54] + PointCLM [55] | | 23.40 | 19.66 | 21.36 |
| CoFiNet [54] + ECC [44] | | 41.49 | 26.60 | 32.41 |
| GeoTransformer [38] + T-Linkage [33] | | 27.68 | 26.40 | 27.02 |
| GeoTransformer [38] + RansaCov [34] | 69.65 | 41.86 | 48.16 | 44.78 |
| GeoTransformer [38] + PointCLM [55] | | 68.09 | 69.09 | 68.58 |
| GeoTransformer [38] + ECC [44] | | 78.03 | 64.08 | 70.37 |
| MIRETR (ours) + T-Linkage [33] | | 29.40 | 28.67 | 29.03 |
| MIRETR (ours) + RansaCov [34] | | 45.63 | 46.85 | 46.23 |
| MIRETR (ours) + PointCLM [55] | 77.13 | 85.63 | 85.81 | 85.71 |
| MIRETR (ours) + ECC [44] | | 92.56 | 82.08 | 87.01 |
| MIRETR (ours, full pipeline) | | 94.95 | 93.94 | 94.44 |
",
+ "image_path": "843107ec916c96881858bfb81362e9325f0099639ec1d5bd81622569773569d0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 309,
+ 221,
+ 542,
+ 281
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 330,
+ 209,
+ 520,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 330,
+ 209,
+ 520,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 330,
+ 209,
+ 520,
+ 219
+ ],
+ "type": "text",
+ "content": "Table 3. Evaluation results on ShapeNet benchmark."
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 309,
+ 221,
+ 542,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 221,
+ 542,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 221,
+ 542,
+ 281
+ ],
+ "type": "table",
+ "html": "| Dataset | Prompt collection | Image Generation | Preference annotation |
| Source | Annotation | Source | Number | Rating | Dimension |
| DiffusionDB [19] | DiffusionDB | × | Diffusion (1) | 1,819,808 | 0 | None |
| AGIQA-1K [24] | DiffusionDB | × | Diffusion (2) | 1,080 | 23,760 | Overall |
| PickScore [6] | Web Application | × | Diffusion (3) | 583,747 | 583,747 | Overall |
| ImageReward [22] | DiffusionDB | × | Auto Regressive; Diffusion (6) | 136,892 | 410,676 | Overall |
| HPS [21] | DiffusionDB | × | Diffusion (1) | 98,807 | 98,807 | Overall |
| HPS v2 [20] | DiffusionDB, COCO | ✓ | GAN; Auto Regressive; Diffusion, COCO (9) | 430,060 | 798,090 | Overall |
| AGIQA-3K [7] | DiffusionDB | × | GAN; Auto Regressive; Diffusion (6) | 2,982 | 125,244 | Overall; Alignment |
| MHP | DiffusionDB, PromptHero, KOLORS, GPT4 | ✓ | GAN; Auto Regressive; Diffusion (9) | 607,541 | 918,315 | Aesthetics, Detail, Alignment, Overall |
",
+ "bbox": [
+ 84,
+ 169,
+ 883,
+ 369
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "range of personalized needs and preferences. To ensure a comprehensive evaluation of text-to-image synthesis outputs, it is crucial to learn and utilize multi-dimensional human preferences.",
+ "bbox": [
+ 75,
+ 383,
+ 470,
+ 445
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "To learn the multi-dimensional human preferences, we propose the Multi-dimensional Human Preference (MHP) dataset. Compared to prior efforts [6, 20-22], the MHP dataset offers significant enhancements in prompts collection, image generation, and preference annotation. For the prompt collection, previous work [6, 21, 22] directly utilizes existing open-source datasets (e.g., Diffusion DB [19]) or datasets collated from the internet [6], overlooking the potential data bias of long-tail distribution. To this end, based on the categories schema of Parti [23], we annotate the collected prompts into 7 category labels (e.g., characters, scenes, objects, animals, etc.). For the underrepresented tail categories, we employ Large Language Models (LLMs) (e.g., GPT-4 [9]) to generate additional prompts. This process results in a balanced prompt collection across various categories, which is used for later image generation. For image generation, following previous work [7, 20], we not only utilize existing open-source Diffusion models and their variants, but also employ GANs and auto-regressive models to generate images. Consequently, we generate a dataset of 607,541 images, which are further used to create 918,315 pairwise comparisons of images for preference annotation. The quantity of image data constitutes the largest dataset of its kind. For the annotation of human preferences, contrary to the single annotation of existing work [6, 20, 21, 24], we consider a broader range of dimensions for human preferences and employ human annotators to label each image pair across four dimensions, including aesthetics, detail quality, semantic alignment, and overall score.",
+ "bbox": [
+ 75,
+ 446,
+ 470,
+ 883
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "To learn human preferences, existing methods employ",
+ "bbox": [
+ 96,
+ 885,
+ 468,
+ 901
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "the pre-trained vision-language models (e.g., CLIP [11], BLIP [8]) to extract features from images and prompts independently, followed by computing the similarity between them. These methods then fine-tune the network utilizing the collected preference data. For learning the multidimensional preferences, a straightforward strategy is to train separate models for different preferences. However, such a simple strategy requires data re-collection and model re-training for the new preference. Moreover, due to the potential bias in single-preference data, a model trained under one preference condition often exhibits diminished performance when evaluated against other preferences. Therefore, we propose the Multi-dimensional Preference Score (MPS), a unified model capable of predicting scores under various preference conditions. Specifically, a certain preference is denoted by a series of descriptive words. For instance, the 'aesthetic' condition is decomposed into words such as 'light', 'color', and 'clarity' to describe the attributes of this condition. These attribute words are used to compute similarities with the prompt, resulting in a similarity matrix that reflects the correspondence between words in the prompt and the specified condition. On the other hand, features from images and text are extracted using a pre-trained vision-language model. Subsequently, two modalities are fused through a multimodal cross-attention layer. The similarity matrix serves as a mask merged into the cross-attention layer, which ensures that the text only related to the condition is attended to by the visual modality. Then the fused features are used to predict the preference scores. We evaluate our MPS model on both the existing human preference datasets (i.e., ImageReward [22] and HPS v2 [20]) and our MHP dataset. The experimental results indicate that our MPS model surpasses existing benchmarks in evaluating both overall and multi-dimensional prefer",
+ "bbox": [
+ 496,
+ 383,
+ 893,
+ 898
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "8019",
+ "bbox": [
+ 482,
+ 944,
+ 514,
+ 955
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "ences, establishing a new state-of-the-art in comparison with related work.",
+ "bbox": [
+ 76,
+ 90,
+ 468,
+ 119
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In summary, our main contributions are as follows:",
+ "bbox": [
+ 96,
+ 122,
+ 433,
+ 136
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- We introduce the Multi-dimensional Human Preference (MHP) datasets for evaluating text-to-image models. The MHP contains balanced prompts and the largest collection of images with multi-dimensional annotations. Based on MHP, we propose a standard test benchmark to evaluate existing text-to-image synthesis models.",
+ "- We propose the MPS model, which learns multidimensional human preferences and evaluates the scores of generated images under different preference conditions.",
+ "- Our MPS exhibits superior performance compared to existing methods across three datasets in predicting the overall preferences and multi-dimensional preferences."
+ ],
+ "bbox": [
+ 76,
+ 137,
+ 467,
+ 333
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2. Related work",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 348,
+ 212,
+ 364
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2.1. Text-to-image Generation and Evaluation",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 373,
+ 431,
+ 388
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "The text-to-image task aims to synthesize realistic images from natural language description (i.e., prompt). Several work attempts to tackle this problem, including GANs [4, 25], auto-regressive [2, 3, 23] and diffusion models [10, 14, 15]. Among the previously mentioned methods, diffusion models gain significant attention for their exceptional performance. These methods are principally divided into two categories: latent-based and pixel-based approaches. The Latent Diffusion Model (LDM) [14] is notable as the first to introduce a latent-based diffusion model, leveraging an auto-encoder to map images into a latent space where the diffusion process is executed. Following this, Stable Diffusion has notably propelled the field forward by open-sourcing SD series [10]. In contrast, DALL-E 2 [13] and Imagen [15] are predicated on pixel-based diffusion models. Besides, the Imagen [15] integrates the large language model T5 XXL to achieve a text-to-image super-resolution diffusion model capable of producing highly realistic images. Current text-to-image models excel in creating high-quality images but often miss aligning with human preferences in real-world applications. For evaluating text-to-image models, several evaluation metrics are proposed to evaluate the quality of generated images, including Inception Score (IS) [16], Fréchet Inception Distance (FID) [5], and CLIP Score [12]. However, these statistical metrics fall short of aligning with human perceptual preferences. Our MPS provides a comprehensive evaluation for text-to-image generations, facilitating the evaluation of alignment with multi-dimensional human preference.",
+ "bbox": [
+ 76,
+ 397,
+ 468,
+ 835
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2.2. Learning human preferences",
+ "text_level": 1,
+ "bbox": [
+ 76,
+ 845,
+ 334,
+ 863
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Currently, several studies attempt to collect and learn human preferences for the evaluation of text-to-image gen",
+ "bbox": [
+ 76,
+ 869,
+ 467,
+ 900
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "image",
+ "img_path": "images/d353a47d6c6d83dc4c757bfcce712d067a429789e9da4cfb07202e28edbb42ec.jpg",
+ "image_caption": [
+ "Figure 2. Prompt collection. The initially collected prompts exhibit a long-tail distribution across various categories. After prompt augmentation with GPT, we obtain a relatively balanced prompt dataset, which contains 66,389 prompts."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 513,
+ 97,
+ 890,
+ 214
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "eration. They use the collected data to fine-tune visual-language models (VLMs) to align with human selections.",
+ "bbox": [
+ 498,
+ 301,
+ 890,
+ 330
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "HPS [21] introduces the HPD dataset of human preference choices. The images are generated solely using the Stable Diffusion model with prompts from Diffusion DB. They train HPS model utilizing the human preference annotations of HPD to align it with human preference. Subsequently, they fine-tune the Stable Diffusion model under the guidance of HPS, leading to better generated images that are more preferred by human users. However, HPS is limited to a single generation model and a relatively small number of images. Furthermore, HPS v2 [20] introduces a larger dataset, employing 8 generative models, including Diffusion models, GANs, and Auto regressive models, and also incorporating captions from the COCO dataset. However, both HPS and HPS v2 primarily focus on overall human preferences, not considering the diversity of human tastes.",
+ "bbox": [
+ 496,
+ 333,
+ 890,
+ 559
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "PickScore [6] proposes a web application designed to collect prompts and human preference annotations from real users. Unlike previous methods adopting prompts from existing datasets (e.g. DiffusionDB), the prompts of PickScore are directly generated by actual users. The dataset of PickScore is sizeable, however, it focuses only on the overall preferences, lacking detailed annotations for multi-dimension preferences.",
+ "bbox": [
+ 496,
+ 561,
+ 890,
+ 681
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "ImageReward [22] employs four types of Diffusion models along with an auto-regression-based model. Their annotation of generated images is more detailed with scoring ranging from 0 to 7. Beyond overall satisfaction, they also consider annotations for alignment and fidelity. However, they merge alignment and fidelity into a single overall score, inadequately capturing the multi-dimensional dimensions of human preferences.",
+ "bbox": [
+ 496,
+ 684,
+ 890,
+ 805
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "AGIQA-1k [24] and AGIQA-3k [7] utilize a range of generative models, including Diffusion, GAN, and Auto regressive models, to produce images. They consider both overall and alignment preferences. However, their dataset size is considerably smaller compared to existing work.",
+ "bbox": [
+ 498,
+ 806,
+ 890,
+ 883
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Our MHP dataset represents an advancement over previ",
+ "bbox": [
+ 519,
+ 885,
+ 890,
+ 900
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "8020",
+ "bbox": [
+ 482,
+ 944,
+ 514,
+ 955
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "table",
+ "img_path": "images/e3c2010a480f778cbf2f178ecddca0795649d8927f8a1339cdb9c9db9a176d42.jpg",
+ "table_caption": [
+ "Table 2. Image generation. The image sources of the MHP dataset consist of the images generated from 9 text-to-image generative models. Note that KOKORS is an internal model for in-house designer platform."
+ ],
+ "table_footnote": [],
+ "table_body": "| Dataset | Prompt collection | Image Generation | Preference annotation |
| Source | Annotation | Source | Number | Rating | Dimension |
| DiffusionDB [19] | DiffusionDB | × | Diffusion (1) | 1,819,808 | 0 | None |
| AGIQA-1K [24] | DiffusionDB | × | Diffusion (2) | 1,080 | 23,760 | Overall |
| PickScore [6] | Web Application | × | Diffusion (3) | 583,747 | 583,747 | Overall |
| ImageReward [22] | DiffusionDB | × | Auto Regressive; Diffusion (6) | 136,892 | 410,676 | Overall |
| HPS [21] | DiffusionDB | × | Diffusion (1) | 98,807 | 98,807 | Overall |
| HPS v2 [20] | DiffusionDB, COCO | ✓ | GAN; Auto Regressive; Diffusion, COCO (9) | 430,060 | 798,090 | Overall |
| AGIQA-3K [7] | DiffusionDB | × | GAN; Auto Regressive; Diffusion (6) | 2,982 | 125,244 | Overall; Alignment |
| MHP | DiffusionDB, PromptHero, KOLORS, GPT4 | ✓ | GAN; Auto Regressive; Diffusion (9) | 607,541 | 918,315 | Aesthetics, Detail, Alignment, Overall |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.385,
+ 0.471,
+ 0.446
+ ],
+ "angle": 0,
+ "content": "range of personalized needs and preferences. To ensure a comprehensive evaluation of text-to-image synthesis outputs, it is crucial to learn and utilize multi-dimensional human preferences."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.447,
+ 0.471,
+ 0.884
+ ],
+ "angle": 0,
+ "content": "To learn the multi-dimensional human preferences, we propose the Multi-dimensional Human Preference (MHP) dataset. Compared to prior efforts [6, 20-22], the MHP dataset offers significant enhancements in prompts collection, image generation, and preference annotation. For the prompt collection, previous work [6, 21, 22] directly utilizes existing open-source datasets (e.g., Diffusion DB [19]) or datasets collated from the internet [6], overlooking the potential data bias of long-tail distribution. To this end, based on the categories schema of Parti [23], we annotate the collected prompts into 7 category labels (e.g., characters, scenes, objects, animals, etc.). For the underrepresented tail categories, we employ Large Language Models (LLMs) (e.g., GPT-4 [9]) to generate additional prompts. This process results in a balanced prompt collection across various categories, which is used for later image generation. For image generation, following previous work [7, 20], we not only utilize existing open-source Diffusion models and their variants, but also employ GANs and auto-regressive models to generate images. Consequently, we generate a dataset of 607,541 images, which are further used to create 918,315 pairwise comparisons of images for preference annotation. The quantity of image data constitutes the largest dataset of its kind. For the annotation of human preferences, contrary to the single annotation of existing work [6, 20, 21, 24], we consider a broader range of dimensions for human preferences and employ human annotators to label each image pair across four dimensions, including aesthetics, detail quality, semantic alignment, and overall score."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.097,
+ 0.886,
+ 0.47,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "To learn human preferences, existing methods employ"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.385,
+ 0.895,
+ 0.899
+ ],
+ "angle": 0,
+ "content": "the pre-trained vision-language models (e.g., CLIP [11], BLIP [8]) to extract features from images and prompts independently, followed by computing the similarity between them. These methods then fine-tune the network utilizing the collected preference data. For learning the multidimensional preferences, a straightforward strategy is to train separate models for different preferences. However, such a simple strategy requires data re-collection and model re-training for the new preference. Moreover, due to the potential bias in single-preference data, a model trained under one preference condition often exhibits diminished performance when evaluated against other preferences. Therefore, we propose the Multi-dimensional Preference Score (MPS), a unified model capable of predicting scores under various preference conditions. Specifically, a certain preference is denoted by a series of descriptive words. For instance, the 'aesthetic' condition is decomposed into words such as 'light', 'color', and 'clarity' to describe the attributes of this condition. These attribute words are used to compute similarities with the prompt, resulting in a similarity matrix that reflects the correspondence between words in the prompt and the specified condition. On the other hand, features from images and text are extracted using a pre-trained vision-language model. Subsequently, two modalities are fused through a multimodal cross-attention layer. The similarity matrix serves as a mask merged into the cross-attention layer, which ensures that the text only related to the condition is attended to by the visual modality. Then the fused features are used to predict the preference scores. We evaluate our MPS model on both the existing human preference datasets (i.e., ImageReward [22] and HPS v2 [20]) and our MHP dataset. The experimental results indicate that our MPS model surpasses existing benchmarks in evaluating both overall and multi-dimensional prefer"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.483,
+ 0.945,
+ 0.516,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "8019"
+ }
+ ],
+ [
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.092,
+ 0.47,
+ 0.12
+ ],
+ "angle": 0,
+ "content": "ences, establishing a new state-of-the-art in comparison with related work."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.097,
+ 0.123,
+ 0.434,
+ 0.137
+ ],
+ "angle": 0,
+ "content": "In summary, our main contributions are as follows:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.078,
+ 0.138,
+ 0.468,
+ 0.228
+ ],
+ "angle": 0,
+ "content": "- We introduce the Multi-dimensional Human Preference (MHP) datasets for evaluating text-to-image models. The MHP contains balanced prompts and the largest collection of images with multi-dimensional annotations. Based on MHP, we propose a standard test benchmark to evaluate existing text-to-image synthesis models."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.078,
+ 0.229,
+ 0.468,
+ 0.287
+ ],
+ "angle": 0,
+ "content": "- We propose the MPS model, which learns multidimensional human preferences and evaluates the scores of generated images under different preference conditions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.078,
+ 0.289,
+ 0.468,
+ 0.334
+ ],
+ "angle": 0,
+ "content": "- Our MPS exhibits superior performance compared to existing methods across three datasets in predicting the overall preferences and multi-dimensional preferences."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.078,
+ 0.138,
+ 0.468,
+ 0.334
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.349,
+ 0.214,
+ 0.365
+ ],
+ "angle": 0,
+ "content": "2. Related work"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.374,
+ 0.433,
+ 0.39
+ ],
+ "angle": 0,
+ "content": "2.1. Text-to-image Generation and Evaluation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.398,
+ 0.47,
+ 0.837
+ ],
+ "angle": 0,
+ "content": "The text-to-image task aims to synthesize realistic images from natural language description (i.e., prompt). Several work attempts to tackle this problem, including GANs [4, 25], auto-regressive [2, 3, 23] and diffusion models [10, 14, 15]. Among the previously mentioned methods, diffusion models gain significant attention for their exceptional performance. These methods are principally divided into two categories: latent-based and pixel-based approaches. The Latent Diffusion Model (LDM) [14] is notable as the first to introduce a latent-based diffusion model, leveraging an auto-encoder to map images into a latent space where the diffusion process is executed. Following this, Stable Diffusion has notably propelled the field forward by open-sourcing SD series [10]. In contrast, DALL-E 2 [13] and Imagen [15] are predicated on pixel-based diffusion models. Besides, the Imagen [15] integrates the large language model T5 XXL to achieve a text-to-image super-resolution diffusion model capable of producing highly realistic images. Current text-to-image models excel in creating high-quality images but often miss aligning with human preferences in real-world applications. For evaluating text-to-image models, several evaluation metrics are proposed to evaluate the quality of generated images, including Inception Score (IS) [16], Fréchet Inception Distance (FID) [5], and CLIP Score [12]. However, these statistical metrics fall short of aligning with human perceptual preferences. Our MPS provides a comprehensive evaluation for text-to-image generations, facilitating the evaluation of alignment with multi-dimensional human preference."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.078,
+ 0.847,
+ 0.336,
+ 0.864
+ ],
+ "angle": 0,
+ "content": "2.2. Learning human preferences"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.871,
+ 0.468,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Currently, several studies attempt to collect and learn human preferences for the evaluation of text-to-image gen"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.514,
+ 0.098,
+ 0.891,
+ 0.215
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.499,
+ 0.225,
+ 0.892,
+ 0.281
+ ],
+ "angle": 0,
+ "content": "Figure 2. Prompt collection. The initially collected prompts exhibit a long-tail distribution across various categories. After prompt augmentation with GPT, we obtain a relatively balanced prompt dataset, which contains 66,389 prompts."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.499,
+ 0.302,
+ 0.891,
+ 0.332
+ ],
+ "angle": 0,
+ "content": "eration. They use the collected data to fine-tune visual-language models (VLMs) to align with human selections."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.334,
+ 0.892,
+ 0.56
+ ],
+ "angle": 0,
+ "content": "HPS [21] introduces the HPD dataset of human preference choices. The images are generated solely using the Stable Diffusion model with prompts from Diffusion DB. They train HPS model utilizing the human preference annotations of HPD to align it with human preference. Subsequently, they fine-tune the Stable Diffusion model under the guidance of HPS, leading to better generated images that are more preferred by human users. However, HPS is limited to a single generation model and a relatively small number of images. Furthermore, HPS v2 [20] introduces a larger dataset, employing 8 generative models, including Diffusion models, GANs, and Auto regressive models, and also incorporating captions from the COCO dataset. However, both HPS and HPS v2 primarily focus on overall human preferences, not considering the diversity of human tastes."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.563,
+ 0.892,
+ 0.683
+ ],
+ "angle": 0,
+ "content": "PickScore [6] proposes a web application designed to collect prompts and human preference annotations from real users. Unlike previous methods adopting prompts from existing datasets (e.g. DiffusionDB), the prompts of PickScore are directly generated by actual users. The dataset of PickScore is sizeable, however, it focuses only on the overall preferences, lacking detailed annotations for multi-dimension preferences."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.685,
+ 0.892,
+ 0.806
+ ],
+ "angle": 0,
+ "content": "ImageReward [22] employs four types of Diffusion models along with an auto-regression-based model. Their annotation of generated images is more detailed with scoring ranging from 0 to 7. Beyond overall satisfaction, they also consider annotations for alignment and fidelity. However, they merge alignment and fidelity into a single overall score, inadequately capturing the multi-dimensional dimensions of human preferences."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.499,
+ 0.808,
+ 0.892,
+ 0.884
+ ],
+ "angle": 0,
+ "content": "AGIQA-1k [24] and AGIQA-3k [7] utilize a range of generative models, including Diffusion, GAN, and Auto regressive models, to produce images. They consider both overall and alignment preferences. However, their dataset size is considerably smaller compared to existing work."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.52,
+ 0.886,
+ 0.891,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "Our MHP dataset represents an advancement over previ"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.483,
+ 0.945,
+ 0.516,
+ 0.956
+ ],
+ "angle": 0,
+ "content": "8020"
+ }
+ ],
+ [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.077,
+ 0.089,
+ 0.471,
+ 0.146
+ ],
+ "angle": 0,
+ "content": "Table 2. Image generation. The image sources of the MHP dataset consist of the images generated from 9 text-to-image generative models. Note that KOKORS is an internal model for in-house designer platform."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.084,
+ 0.156,
+ 0.462,
+ 0.312
+ ],
+ "angle": 0,
+ "content": "| Dataset | Prompt collection | Image Generation | Preference annotation |
| Source | Annotation | Source | Number | Rating | Dimension |
| DiffusionDB [19] | DiffusionDB | × | Diffusion (1) | 1,819,808 | 0 | None |
| AGIQA-1K [24] | DiffusionDB | × | Diffusion (2) | 1,080 | 23,760 | Overall |
| PickScore [6] | Web Application | × | Diffusion (3) | 583,747 | 583,747 | Overall |
| ImageReward [22] | DiffusionDB | × | Auto Regressive; Diffusion (6) | 136,892 | 410,676 | Overall |
| HPS [21] | DiffusionDB | × | Diffusion (1) | 98,807 | 98,807 | Overall |
| HPS v2 [20] | DiffusionDB, COCO | ✓ | GAN; Auto Regressive; Diffusion, COCO (9) | 430,060 | 798,090 | Overall |
| AGIQA-3K [7] | DiffusionDB | × | GAN; Auto Regressive; Diffusion (6) | 2,982 | 125,244 | Overall; Alignment |
| MHP | DiffusionDB, PromptHero, KOLORS, GPT4 | ✓ | GAN; Auto Regressive; Diffusion (9) | 607,541 | 918,315 | Aesthetics, Detail, Alignment, Overall |
+
+range of personalized needs and preferences. To ensure a comprehensive evaluation of text-to-image synthesis outputs, it is crucial to learn and utilize multi-dimensional human preferences.
+
+To learn the multi-dimensional human preferences, we propose the Multi-dimensional Human Preference (MHP) dataset. Compared to prior efforts [6, 20-22], the MHP dataset offers significant enhancements in prompts collection, image generation, and preference annotation. For the prompt collection, previous work [6, 21, 22] directly utilizes existing open-source datasets (e.g., Diffusion DB [19]) or datasets collated from the internet [6], overlooking the potential data bias of long-tail distribution. To this end, based on the categories schema of Parti [23], we annotate the collected prompts into 7 category labels (e.g., characters, scenes, objects, animals, etc.). For the underrepresented tail categories, we employ Large Language Models (LLMs) (e.g., GPT-4 [9]) to generate additional prompts. This process results in a balanced prompt collection across various categories, which is used for later image generation. For image generation, following previous work [7, 20], we not only utilize existing open-source Diffusion models and their variants, but also employ GANs and auto-regressive models to generate images. Consequently, we generate a dataset of 607,541 images, which are further used to create 918,315 pairwise comparisons of images for preference annotation. The quantity of image data constitutes the largest dataset of its kind. For the annotation of human preferences, contrary to the single annotation of existing work [6, 20, 21, 24], we consider a broader range of dimensions for human preferences and employ human annotators to label each image pair across four dimensions, including aesthetics, detail quality, semantic alignment, and overall score.
+
+To learn human preferences, existing methods employ
+
+the pre-trained vision-language models (e.g., CLIP [11], BLIP [8]) to extract features from images and prompts independently, followed by computing the similarity between them. These methods then fine-tune the network utilizing the collected preference data. For learning the multidimensional preferences, a straightforward strategy is to train separate models for different preferences. However, such a simple strategy requires data re-collection and model re-training for the new preference. Moreover, due to the potential bias in single-preference data, a model trained under one preference condition often exhibits diminished performance when evaluated against other preferences. Therefore, we propose the Multi-dimensional Preference Score (MPS), a unified model capable of predicting scores under various preference conditions. Specifically, a certain preference is denoted by a series of descriptive words. For instance, the 'aesthetic' condition is decomposed into words such as 'light', 'color', and 'clarity' to describe the attributes of this condition. These attribute words are used to compute similarities with the prompt, resulting in a similarity matrix that reflects the correspondence between words in the prompt and the specified condition. On the other hand, features from images and text are extracted using a pre-trained vision-language model. Subsequently, two modalities are fused through a multimodal cross-attention layer. The similarity matrix serves as a mask merged into the cross-attention layer, which ensures that the text only related to the condition is attended to by the visual modality. Then the fused features are used to predict the preference scores. We evaluate our MPS model on both the existing human preference datasets (i.e., ImageReward [22] and HPS v2 [20]) and our MHP dataset. The experimental results indicate that our MPS model surpasses existing benchmarks in evaluating both overall and multi-dimensional prefer
+
+ences, establishing a new state-of-the-art in comparison with related work.
+
+In summary, our main contributions are as follows:
+
+- We introduce the Multi-dimensional Human Preference (MHP) datasets for evaluating text-to-image models. The MHP contains balanced prompts and the largest collection of images with multi-dimensional annotations. Based on MHP, we propose a standard test benchmark to evaluate existing text-to-image synthesis models.
+- We propose the MPS model, which learns multidimensional human preferences and evaluates the scores of generated images under different preference conditions.
+- Our MPS exhibits superior performance compared to existing methods across three datasets in predicting the overall preferences and multi-dimensional preferences.
+
+# 2. Related work
+
+# 2.1. Text-to-image Generation and Evaluation
+
+The text-to-image task aims to synthesize realistic images from natural language description (i.e., prompt). Several work attempts to tackle this problem, including GANs [4, 25], auto-regressive [2, 3, 23] and diffusion models [10, 14, 15]. Among the previously mentioned methods, diffusion models gain significant attention for their exceptional performance. These methods are principally divided into two categories: latent-based and pixel-based approaches. The Latent Diffusion Model (LDM) [14] is notable as the first to introduce a latent-based diffusion model, leveraging an auto-encoder to map images into a latent space where the diffusion process is executed. Following this, Stable Diffusion has notably propelled the field forward by open-sourcing SD series [10]. In contrast, DALL-E 2 [13] and Imagen [15] are predicated on pixel-based diffusion models. Besides, the Imagen [15] integrates the large language model T5 XXL to achieve a text-to-image super-resolution diffusion model capable of producing highly realistic images. Current text-to-image models excel in creating high-quality images but often miss aligning with human preferences in real-world applications. For evaluating text-to-image models, several evaluation metrics are proposed to evaluate the quality of generated images, including Inception Score (IS) [16], Fréchet Inception Distance (FID) [5], and CLIP Score [12]. However, these statistical metrics fall short of aligning with human perceptual preferences. Our MPS provides a comprehensive evaluation for text-to-image generations, facilitating the evaluation of alignment with multi-dimensional human preference.
+
+# 2.2. Learning human preferences
+
+Currently, several studies attempt to collect and learn human preferences for the evaluation of text-to-image gen
+
+
+Figure 2. Prompt collection. The initially collected prompts exhibit a long-tail distribution across various categories. After prompt augmentation with GPT, we obtain a relatively balanced prompt dataset, which contains 66,389 prompts.
+
+eration. They use the collected data to fine-tune visual-language models (VLMs) to align with human selections.
+
+HPS [21] introduces the HPD dataset of human preference choices. The images are generated solely using the Stable Diffusion model with prompts from Diffusion DB. They train HPS model utilizing the human preference annotations of HPD to align it with human preference. Subsequently, they fine-tune the Stable Diffusion model under the guidance of HPS, leading to better generated images that are more preferred by human users. However, HPS is limited to a single generation model and a relatively small number of images. Furthermore, HPS v2 [20] introduces a larger dataset, employing 8 generative models, including Diffusion models, GANs, and Auto regressive models, and also incorporating captions from the COCO dataset. However, both HPS and HPS v2 primarily focus on overall human preferences, not considering the diversity of human tastes.
+
+PickScore [6] proposes a web application designed to collect prompts and human preference annotations from real users. Unlike previous methods adopting prompts from existing datasets (e.g. DiffusionDB), the prompts of PickScore are directly generated by actual users. The dataset of PickScore is sizeable, however, it focuses only on the overall preferences, lacking detailed annotations for multi-dimension preferences.
+
+ImageReward [22] employs four types of Diffusion models along with an auto-regression-based model. Their annotation of generated images is more detailed with scoring ranging from 0 to 7. Beyond overall satisfaction, they also consider annotations for alignment and fidelity. However, they merge alignment and fidelity into a single overall score, inadequately capturing the multi-dimensional dimensions of human preferences.
+
+AGIQA-1k [24] and AGIQA-3k [7] utilize a range of generative models, including Diffusion, GAN, and Auto regressive models, to produce images. They consider both overall and alignment preferences. However, their dataset size is considerably smaller compared to existing work.
+
+Our MHP dataset represents an advancement over previ
+
+Table 2. Image generation. The image sources of the MHP dataset consist of the images generated from 9 text-to-image generative models. Note that KOKORS is an internal model for in-house designer platform.
+
+| Dataset | Prompt collection | Image Generation | Preference annotation |
| Source | Annotation | Source | Number | Rating | Dimension |
| DiffusionDB [19] | DiffusionDB | × | Diffusion (1) | 1,819,808 | 0 | None |
| AGIQA-1K [24] | DiffusionDB | × | Diffusion (2) | 1,080 | 23,760 | Overall |
| PickScore [6] | Web Application | × | Diffusion (3) | 583,747 | 583,747 | Overall |
| ImageReward [22] | DiffusionDB | × | Auto Regressive; Diffusion (6) | 136,892 | 410,676 | Overall |
| HPS [21] | DiffusionDB | × | Diffusion (1) | 98,807 | 98,807 | Overall |
| HPS v2 [20] | DiffusionDB, COCO | ✓ | GAN; Auto Regressive; Diffusion, COCO (9) | 430,060 | 798,090 | Overall |
| AGIQA-3K [7] | DiffusionDB | × | GAN; Auto Regressive; Diffusion (6) | 2,982 | 125,244 | Overall; Alignment |
| MHP | DiffusionDB, PromptHero, KOLORS, GPT4 | ✓ | GAN; Auto Regressive; Diffusion (9) | 607,541 | 918,315 | Aesthetics, Detail, Alignment, Overall |
",
+ "image_path": "17eb6dcfb4561a35d90b4a96dacb901987ffbcdcb5739290e4ebdc54cdea9d63.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 46,
+ 304,
+ 288,
+ 353
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 304,
+ 288,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 304,
+ 288,
+ 353
+ ],
+ "type": "text",
+ "content": "range of personalized needs and preferences. To ensure a comprehensive evaluation of text-to-image synthesis outputs, it is crucial to learn and utilize multi-dimensional human preferences."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 46,
+ 354,
+ 288,
+ 700
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 354,
+ 288,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 354,
+ 288,
+ 700
+ ],
+ "type": "text",
+ "content": "To learn the multi-dimensional human preferences, we propose the Multi-dimensional Human Preference (MHP) dataset. Compared to prior efforts [6, 20-22], the MHP dataset offers significant enhancements in prompts collection, image generation, and preference annotation. For the prompt collection, previous work [6, 21, 22] directly utilizes existing open-source datasets (e.g., Diffusion DB [19]) or datasets collated from the internet [6], overlooking the potential data bias of long-tail distribution. To this end, based on the categories schema of Parti [23], we annotate the collected prompts into 7 category labels (e.g., characters, scenes, objects, animals, etc.). For the underrepresented tail categories, we employ Large Language Models (LLMs) (e.g., GPT-4 [9]) to generate additional prompts. This process results in a balanced prompt collection across various categories, which is used for later image generation. For image generation, following previous work [7, 20], we not only utilize existing open-source Diffusion models and their variants, but also employ GANs and auto-regressive models to generate images. Consequently, we generate a dataset of 607,541 images, which are further used to create 918,315 pairwise comparisons of images for preference annotation. The quantity of image data constitutes the largest dataset of its kind. For the annotation of human preferences, contrary to the single annotation of existing work [6, 20, 21, 24], we consider a broader range of dimensions for human preferences and employ human annotators to label each image pair across four dimensions, including aesthetics, detail quality, semantic alignment, and overall score."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 59,
+ 701,
+ 287,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 59,
+ 701,
+ 287,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 59,
+ 701,
+ 287,
+ 714
+ ],
+ "type": "text",
+ "content": "To learn human preferences, existing methods employ"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 304,
+ 304,
+ 547,
+ 712
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 304,
+ 547,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 304,
+ 547,
+ 712
+ ],
+ "type": "text",
+ "content": "the pre-trained vision-language models (e.g., CLIP [11], BLIP [8]) to extract features from images and prompts independently, followed by computing the similarity between them. These methods then fine-tune the network utilizing the collected preference data. For learning the multidimensional preferences, a straightforward strategy is to train separate models for different preferences. However, such a simple strategy requires data re-collection and model re-training for the new preference. Moreover, due to the potential bias in single-preference data, a model trained under one preference condition often exhibits diminished performance when evaluated against other preferences. Therefore, we propose the Multi-dimensional Preference Score (MPS), a unified model capable of predicting scores under various preference conditions. Specifically, a certain preference is denoted by a series of descriptive words. For instance, the 'aesthetic' condition is decomposed into words such as 'light', 'color', and 'clarity' to describe the attributes of this condition. These attribute words are used to compute similarities with the prompt, resulting in a similarity matrix that reflects the correspondence between words in the prompt and the specified condition. On the other hand, features from images and text are extracted using a pre-trained vision-language model. Subsequently, two modalities are fused through a multimodal cross-attention layer. The similarity matrix serves as a mask merged into the cross-attention layer, which ensures that the text only related to the condition is attended to by the visual modality. Then the fused features are used to predict the preference scores. We evaluate our MPS model on both the existing human preference datasets (i.e., ImageReward [22] and HPS v2 [20]) and our MHP dataset. The experimental results indicate that our MPS model surpasses existing benchmarks in evaluating both overall and multi-dimensional prefer"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "text",
+ "content": "8019"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 47,
+ 72,
+ 287,
+ 95
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 72,
+ 287,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 72,
+ 287,
+ 95
+ ],
+ "type": "text",
+ "content": "ences, establishing a new state-of-the-art in comparison with related work."
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 59,
+ 97,
+ 265,
+ 108
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 59,
+ 97,
+ 265,
+ 108
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 59,
+ 97,
+ 265,
+ 108
+ ],
+ "type": "text",
+ "content": "In summary, our main contributions are as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 47,
+ 109,
+ 286,
+ 264
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 5,
+ "blocks": [
+ {
+ "bbox": [
+ 47,
+ 109,
+ 286,
+ 180
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 109,
+ 286,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 109,
+ 286,
+ 180
+ ],
+ "type": "text",
+ "content": "- We introduce the Multi-dimensional Human Preference (MHP) datasets for evaluating text-to-image models. The MHP contains balanced prompts and the largest collection of images with multi-dimensional annotations. Based on MHP, we propose a standard test benchmark to evaluate existing text-to-image synthesis models."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 47,
+ 181,
+ 286,
+ 227
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 181,
+ 286,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 181,
+ 286,
+ 227
+ ],
+ "type": "text",
+ "content": "- We propose the MPS model, which learns multidimensional human preferences and evaluates the scores of generated images under different preference conditions."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 47,
+ 228,
+ 286,
+ 264
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 228,
+ 286,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 228,
+ 286,
+ 264
+ ],
+ "type": "text",
+ "content": "- Our MPS exhibits superior performance compared to existing methods across three datasets in predicting the overall preferences and multi-dimensional preferences."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 47,
+ 276,
+ 130,
+ 289
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 276,
+ 130,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 276,
+ 130,
+ 289
+ ],
+ "type": "text",
+ "content": "2. Related work"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 47,
+ 296,
+ 264,
+ 308
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 296,
+ 264,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 296,
+ 264,
+ 308
+ ],
+ "type": "text",
+ "content": "2.1. Text-to-image Generation and Evaluation"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 47,
+ 315,
+ 287,
+ 662
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 315,
+ 287,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 315,
+ 287,
+ 662
+ ],
+ "type": "text",
+ "content": "The text-to-image task aims to synthesize realistic images from natural language description (i.e., prompt). Several work attempts to tackle this problem, including GANs [4, 25], auto-regressive [2, 3, 23] and diffusion models [10, 14, 15]. Among the previously mentioned methods, diffusion models gain significant attention for their exceptional performance. These methods are principally divided into two categories: latent-based and pixel-based approaches. The Latent Diffusion Model (LDM) [14] is notable as the first to introduce a latent-based diffusion model, leveraging an auto-encoder to map images into a latent space where the diffusion process is executed. Following this, Stable Diffusion has notably propelled the field forward by open-sourcing SD series [10]. In contrast, DALL-E 2 [13] and Imagen [15] are predicated on pixel-based diffusion models. Besides, the Imagen [15] integrates the large language model T5 XXL to achieve a text-to-image super-resolution diffusion model capable of producing highly realistic images. Current text-to-image models excel in creating high-quality images but often miss aligning with human preferences in real-world applications. For evaluating text-to-image models, several evaluation metrics are proposed to evaluate the quality of generated images, including Inception Score (IS) [16], Fréchet Inception Distance (FID) [5], and CLIP Score [12]. However, these statistical metrics fall short of aligning with human perceptual preferences. Our MPS provides a comprehensive evaluation for text-to-image generations, facilitating the evaluation of alignment with multi-dimensional human preference."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 47,
+ 670,
+ 205,
+ 684
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 670,
+ 205,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 670,
+ 205,
+ 684
+ ],
+ "type": "text",
+ "content": "2.2. Learning human preferences"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 47,
+ 689,
+ 286,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 689,
+ 286,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 689,
+ 286,
+ 713
+ ],
+ "type": "text",
+ "content": "Currently, several studies attempt to collect and learn human preferences for the evaluation of text-to-image gen"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 314,
+ 77,
+ 545,
+ 170
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 314,
+ 77,
+ 545,
+ 170
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 314,
+ 77,
+ 545,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 314,
+ 77,
+ 545,
+ 170
+ ],
+ "type": "image",
+ "image_path": "d353a47d6c6d83dc4c757bfcce712d067a429789e9da4cfb07202e28edbb42ec.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 305,
+ 178,
+ 545,
+ 222
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 178,
+ 545,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 178,
+ 545,
+ 222
+ ],
+ "type": "text",
+ "content": "Figure 2. Prompt collection. The initially collected prompts exhibit a long-tail distribution across various categories. After prompt augmentation with GPT, we obtain a relatively balanced prompt dataset, which contains 66,389 prompts."
+ }
+ ]
+ }
+ ],
+ "index": 12,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 305,
+ 239,
+ 545,
+ 262
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 239,
+ 545,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 239,
+ 545,
+ 262
+ ],
+ "type": "text",
+ "content": "eration. They use the collected data to fine-tune visual-language models (VLMs) to align with human selections."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 304,
+ 264,
+ 545,
+ 443
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 264,
+ 545,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 264,
+ 545,
+ 443
+ ],
+ "type": "text",
+ "content": "HPS [21] introduces the HPD dataset of human preference choices. The images are generated solely using the Stable Diffusion model with prompts from Diffusion DB. They train HPS model utilizing the human preference annotations of HPD to align it with human preference. Subsequently, they fine-tune the Stable Diffusion model under the guidance of HPS, leading to better generated images that are more preferred by human users. However, HPS is limited to a single generation model and a relatively small number of images. Furthermore, HPS v2 [20] introduces a larger dataset, employing 8 generative models, including Diffusion models, GANs, and Auto regressive models, and also incorporating captions from the COCO dataset. However, both HPS and HPS v2 primarily focus on overall human preferences, not considering the diversity of human tastes."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 304,
+ 445,
+ 545,
+ 540
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 445,
+ 545,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 445,
+ 545,
+ 540
+ ],
+ "type": "text",
+ "content": "PickScore [6] proposes a web application designed to collect prompts and human preference annotations from real users. Unlike previous methods adopting prompts from existing datasets (e.g. DiffusionDB), the prompts of PickScore are directly generated by actual users. The dataset of PickScore is sizeable, however, it focuses only on the overall preferences, lacking detailed annotations for multi-dimension preferences."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 304,
+ 542,
+ 545,
+ 638
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 545,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 542,
+ 545,
+ 638
+ ],
+ "type": "text",
+ "content": "ImageReward [22] employs four types of Diffusion models along with an auto-regression-based model. Their annotation of generated images is more detailed with scoring ranging from 0 to 7. Beyond overall satisfaction, they also consider annotations for alignment and fidelity. However, they merge alignment and fidelity into a single overall score, inadequately capturing the multi-dimensional dimensions of human preferences."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 305,
+ 639,
+ 545,
+ 700
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 639,
+ 545,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 639,
+ 545,
+ 700
+ ],
+ "type": "text",
+ "content": "AGIQA-1k [24] and AGIQA-3k [7] utilize a range of generative models, including Diffusion, GAN, and Auto regressive models, to produce images. They consider both overall and alignment preferences. However, their dataset size is considerably smaller compared to existing work."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 318,
+ 701,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 701,
+ 545,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 701,
+ 545,
+ 713
+ ],
+ "type": "text",
+ "content": "Our MHP dataset represents an advancement over previ"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 295,
+ 748,
+ 315,
+ 757
+ ],
+ "type": "text",
+ "content": "8020"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 51,
+ 123,
+ 282,
+ 247
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 47,
+ 70,
+ 288,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 70,
+ 288,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 70,
+ 288,
+ 115
+ ],
+ "type": "text",
+ "content": "Table 2. Image generation. The image sources of the MHP dataset consist of the images generated from 9 text-to-image generative models. Note that KOKORS is an internal model for in-house designer platform."
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 51,
+ 123,
+ 282,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 51,
+ 123,
+ 282,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 51,
+ 123,
+ 282,
+ 247
+ ],
+ "type": "table",
+ "html": " | Without data augmentation | With data augmentation | Colmap |
| Inference | Inference + BA | Inference | Inference + BA |
| Ours | DPESFM | Ours | DPESFM | Ours | DPESFM | Ours | DPESFM |
| Reprojection error (px) | Alcatraz Courtyard | 68.41 | 68.50 | 1.70 | 0.81 (0.82) | 36.01 | 92.37 | 0.81 | 0.92 | 0.81 |
| Alcatraz Water Tower | 36.30 | 50.47 | 1.05 | 1.13 (0.55) | 87.67 | 2831.94 | 0.88 | 10.16 | 0.55 |
| Drinking Fountain Somewhere in Zurich | 36.02 | 45.87 | 0.55 | 0.31 (7.21) | 219.75 | 234.90 | 0.31 | 6.73 | 0.31 |
| Nijo Castle Gate | 65.71 | 64.53 | 0.73 | 0.73 (5.81) | 61.41 | 68.19 | 0.88 | 0.89 | 0.73 |
| Porta San Donato Bologna | 74.32 | 94.26 | 0.74 | 0.74 (1.10) | 52.15 | 84.46 | 0.76 | 0.75 | 0.75 |
| Round Church Cambridge | 61.18 | 55.51 | 0.39 | 0.39 (0.50) | 29.80 | 59.54 | 0.39 | 1.49 | 0.39 |
| Smolny Cathedral St Petersburg | 156.16 | 120.85 | 0.81 | 0.81 (15.15) | 85.38 | 87.81 | 0.81 | 0.81 | 0.81 |
| Some Cathedral in Barcelona | 150.01 | 146.10 | 10.42 | 12.71 (21.46) | 125.68 | 687.83 | 0.89 | 16.77 | 0.89 |
| Sri Veeramakaliamman Singapore | 121.02 | 157.39 | 2.19 | 16.87 (16.92) | 83.50 | 166.68 | 2.13 | 9.30 | 0.71 |
| Yueh Hai Ching Temple Singapore | 37.46 | 52.59 | 0.65 | 0.65 (1.16) | 25.60 | 51.35 | 0.65 | 0.73 | 0.65 |
| Average | 80.66 | 85.61 | 1.92 | 3.52 (7.07) | 80.69 | 436.51 | 0.85 | 4.86 | 0.66 |
| Rotation error (deg) | Alcatraz Courtyard | 10.102 | 13.201 | 2.607 | 0.035 | 6.093 | 10.946 | 0.038 | 0.030 | 0.043 |
| Alcatraz Water Tower | 10.637 | 11.053 | 0.499 | 0.764 | 11.501 | 10.641 | 0.699 | 19.351 | 0.228 |
| Drinking Fountain Somewhere in Zurich | 15.846 | 16.014 | 0.003 | 0.001 | 15.415 | 15.704 | 0.001 | 22.759 | 0.007 |
| Nijo Castle Gate | 16.751 | 10.546 | 0.062 | 0.062 | 17.347 | 20.032 | 0.038 | 0.036 | 0.064 |
| Porta San Donato Bologna | 23.839 | 24.120 | 0.095 | 0.094 | 18.411 | 25.004 | 0.094 | 0.094 | 0.099 |
| Round Church Cambridge | 18.906 | 14.473 | 0.029 | 0.026 | 10.295 | 18.685 | 0.029 | 1.086 | 0.035 |
| Smolny Cathedral St Petersburg | 19.387 | 17.971 | 0.023 | 0.022 | 11.662 | 14.380 | 0.023 | 0.019 | 0.029 |
| Some Cathedral in Barcelona | 27.270 | 30.471 | 10.009 | 20.050 | 27.908 | 29.119 | 0.020 | 47.892 | 0.025 |
| Sri Veeramakaliamman Singapore | 28.275 | 36.903 | 0.549 | 4.871 | 23.702 | 36.176 | 0.457 | 2.759 | 0.169 |
| Yueh Hai Ching Temple Singapore | 15.733 | 22.706 | 0.038 | 0.038 | 9.515 | 21.561 | 0.038 | 0.038 | 0.043 |
| Average | 18.675 | 19.746 | 1.391 | 2.596 | 15.185 | 20.225 | 0.144 | 9.406 | 0.074 |
| Translation error (m) | Alcatraz Courtyard | 4.82 | 4.93 | 1.09 | 0.01 | 2.73 | 5.74 | 0.01 | 0.01 | 0.01 |
| Alcatraz Water Tower | 8.66 | 7.35 | 0.31 | 0.44 | 7.53 | 7.77 | 0.41 | 9.05 | 0.12 |
| Drinking Fountain Somewhere in Zurich | 4.44 | 4.44 | 0.00 | 0.00 | 4.45 | 4.46 | 0.00 | 1.38 | 0.00 |
| Nijo Castle Gate | 5.07 | 3.08 | 0.01 | 0.01 | 5.67 | 6.95 | 0.01 | 0.01 | 0.01 |
| Porta San Donato Bologna | 9.50 | 10.72 | 0.05 | 0.05 | 4.80 | 10.48 | 0.05 | 0.05 | 0.05 |
| Round Church Cambridge | 8.94 | 7.19 | 0.01 | 0.01 | 5.28 | 9.00 | 0.01 | 0.56 | 0.01 |
| Smolny Cathedral St Petersburg | 2.70 | 2.43 | 0.01 | 0.01 | 2.33 | 2.52 | 0.01 | 0.01 | 0.01 |
| Some Cathedral in Barcelona | 12.64 | 12.69 | 3.22 | 6.38 | 12.32 | 12.66 | 0.01 | 11.93 | 0.01 |
| Sri Veeramakaliamman Singapore | 4.94 | 4.90 | 0.16 | 1.32 | 4.93 | 4.90 | 0.14 | 0.77 | 0.04 |
| Yueh Hai Ching Temple Singapore | 4.12 | 4.27 | 0.01 | 0.01 | 2.44 | 4.28 | 0.01 | 0.01 | 0.01 |
| Average | 6.58 | 6.20 | 0.49 | 0.82 | 5.25 | 6.88 | 0.07 | 2.38 | 0.03 |
",
+ "bbox": [
+ 76,
+ 87,
+ 890,
+ 506
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "table",
+ "img_path": "images/73289d5b410e5b30c4303de0d567afe882d8588c07b82357653bb84ed13379a1.jpg",
+ "table_caption": [
+ "Table 1. Results of Euclidean reconstruction of novel test scenes, with and without data augmentation. The results of DPESFM [31] have been acquired by us training the model, along with the results reported by [31] in parentheses, if available. The result of Colmap, as reported by [31], is also added for reference."
+ ],
+ "table_footnote": [],
+ "table_body": " | Without data augmentation | With data augmentation | Colmap |
| Inference | Inference + BA | Inference | Inference + BA |
| Ours | DPESFM | Ours | DPESFM | Ours | DPESFM | Ours | DPESFM |
| Reprojection error (px) | Alcatraz Courtyard | 68.41 | 68.50 | 1.70 | 0.81 (0.82) | 36.01 | 92.37 | 0.81 | 0.92 | 0.81 |
| Alcatraz Water Tower | 36.30 | 50.47 | 1.05 | 1.13 (0.55) | 87.67 | 2831.94 | 0.88 | 10.16 | 0.55 |
| Drinking Fountain Somewhere in Zurich | 36.02 | 45.87 | 0.55 | 0.31 (7.21) | 219.75 | 234.90 | 0.31 | 6.73 | 0.31 |
| Nijo Castle Gate | 65.71 | 64.53 | 0.73 | 0.73 (5.81) | 61.41 | 68.19 | 0.88 | 0.89 | 0.73 |
| Porta San Donato Bologna | 74.32 | 94.26 | 0.74 | 0.74 (1.10) | 52.15 | 84.46 | 0.76 | 0.75 | 0.75 |
| Round Church Cambridge | 61.18 | 55.51 | 0.39 | 0.39 (0.50) | 29.80 | 59.54 | 0.39 | 1.49 | 0.39 |
| Smolny Cathedral St Petersburg | 156.16 | 120.85 | 0.81 | 0.81 (15.15) | 85.38 | 87.81 | 0.81 | 0.81 | 0.81 |
| Some Cathedral in Barcelona | 150.01 | 146.10 | 10.42 | 12.71 (21.46) | 125.68 | 687.83 | 0.89 | 16.77 | 0.89 |
| Sri Veeramakaliamman Singapore | 121.02 | 157.39 | 2.19 | 16.87 (16.92) | 83.50 | 166.68 | 2.13 | 9.30 | 0.71 |
| Yueh Hai Ching Temple Singapore | 37.46 | 52.59 | 0.65 | 0.65 (1.16) | 25.60 | 51.35 | 0.65 | 0.73 | 0.65 |
| Average | 80.66 | 85.61 | 1.92 | 3.52 (7.07) | 80.69 | 436.51 | 0.85 | 4.86 | 0.66 |
| Rotation error (deg) | Alcatraz Courtyard | 10.102 | 13.201 | 2.607 | 0.035 | 6.093 | 10.946 | 0.038 | 0.030 | 0.043 |
| Alcatraz Water Tower | 10.637 | 11.053 | 0.499 | 0.764 | 11.501 | 10.641 | 0.699 | 19.351 | 0.228 |
| Drinking Fountain Somewhere in Zurich | 15.846 | 16.014 | 0.003 | 0.001 | 15.415 | 15.704 | 0.001 | 22.759 | 0.007 |
| Nijo Castle Gate | 16.751 | 10.546 | 0.062 | 0.062 | 17.347 | 20.032 | 0.038 | 0.036 | 0.064 |
| Porta San Donato Bologna | 23.839 | 24.120 | 0.095 | 0.094 | 18.411 | 25.004 | 0.094 | 0.094 | 0.099 |
| Round Church Cambridge | 18.906 | 14.473 | 0.029 | 0.026 | 10.295 | 18.685 | 0.029 | 1.086 | 0.035 |
| Smolny Cathedral St Petersburg | 19.387 | 17.971 | 0.023 | 0.022 | 11.662 | 14.380 | 0.023 | 0.019 | 0.029 |
| Some Cathedral in Barcelona | 27.270 | 30.471 | 10.009 | 20.050 | 27.908 | 29.119 | 0.020 | 47.892 | 0.025 |
| Sri Veeramakaliamman Singapore | 28.275 | 36.903 | 0.549 | 4.871 | 23.702 | 36.176 | 0.457 | 2.759 | 0.169 |
| Yueh Hai Ching Temple Singapore | 15.733 | 22.706 | 0.038 | 0.038 | 9.515 | 21.561 | 0.038 | 0.038 | 0.043 |
| Average | 18.675 | 19.746 | 1.391 | 2.596 | 15.185 | 20.225 | 0.144 | 9.406 | 0.074 |
| Translation error (m) | Alcatraz Courtyard | 4.82 | 4.93 | 1.09 | 0.01 | 2.73 | 5.74 | 0.01 | 0.01 | 0.01 |
| Alcatraz Water Tower | 8.66 | 7.35 | 0.31 | 0.44 | 7.53 | 7.77 | 0.41 | 9.05 | 0.12 |
| Drinking Fountain Somewhere in Zurich | 4.44 | 4.44 | 0.00 | 0.00 | 4.45 | 4.46 | 0.00 | 1.38 | 0.00 |
| Nijo Castle Gate | 5.07 | 3.08 | 0.01 | 0.01 | 5.67 | 6.95 | 0.01 | 0.01 | 0.01 |
| Porta San Donato Bologna | 9.50 | 10.72 | 0.05 | 0.05 | 4.80 | 10.48 | 0.05 | 0.05 | 0.05 |
| Round Church Cambridge | 8.94 | 7.19 | 0.01 | 0.01 | 5.28 | 9.00 | 0.01 | 0.56 | 0.01 |
| Smolny Cathedral St Petersburg | 2.70 | 2.43 | 0.01 | 0.01 | 2.33 | 2.52 | 0.01 | 0.01 | 0.01 |
| Some Cathedral in Barcelona | 12.64 | 12.69 | 3.22 | 6.38 | 12.32 | 12.66 | 0.01 | 11.93 | 0.01 |
| Sri Veeramakaliamman Singapore | 4.94 | 4.90 | 0.16 | 1.32 | 4.93 | 4.90 | 0.14 | 0.77 | 0.04 |
| Yueh Hai Ching Temple Singapore | 4.12 | 4.27 | 0.01 | 0.01 | 2.44 | 4.28 | 0.01 | 0.01 | 0.01 |
| Average | 6.58 | 6.20 | 0.49 | 0.82 | 5.25 | 6.88 | 0.07 | 2.38 | 0.03 |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.076,
+ 0.518,
+ 0.893,
+ 0.561
+ ],
+ "angle": 0,
+ "content": "Table 1. Results of Euclidean reconstruction of novel test scenes, with and without data augmentation. The results of DPESFM [31] have been acquired by us training the model, along with the results reported by [31] in parentheses, if available. The result of Colmap, as reported by [31], is also added for reference."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.085,
+ 0.583,
+ 0.472,
+ 0.719
+ ],
+ "angle": 0,
+ "content": " | Without data augmentation | With data augmentation | Colmap |
| Inference | Inference + BA | Inference | Inference + BA |
| Ours | DPESFM | Ours | DPESFM | Ours | DPESFM | Ours | DPESFM |
| Reprojection error (px) | Alcatraz Courtyard | 68.41 | 68.50 | 1.70 | 0.81 (0.82) | 36.01 | 92.37 | 0.81 | 0.92 | 0.81 |
| Alcatraz Water Tower | 36.30 | 50.47 | 1.05 | 1.13 (0.55) | 87.67 | 2831.94 | 0.88 | 10.16 | 0.55 |
| Drinking Fountain Somewhere in Zurich | 36.02 | 45.87 | 0.55 | 0.31 (7.21) | 219.75 | 234.90 | 0.31 | 6.73 | 0.31 |
| Nijo Castle Gate | 65.71 | 64.53 | 0.73 | 0.73 (5.81) | 61.41 | 68.19 | 0.88 | 0.89 | 0.73 |
| Porta San Donato Bologna | 74.32 | 94.26 | 0.74 | 0.74 (1.10) | 52.15 | 84.46 | 0.76 | 0.75 | 0.75 |
| Round Church Cambridge | 61.18 | 55.51 | 0.39 | 0.39 (0.50) | 29.80 | 59.54 | 0.39 | 1.49 | 0.39 |
| Smolny Cathedral St Petersburg | 156.16 | 120.85 | 0.81 | 0.81 (15.15) | 85.38 | 87.81 | 0.81 | 0.81 | 0.81 |
| Some Cathedral in Barcelona | 150.01 | 146.10 | 10.42 | 12.71 (21.46) | 125.68 | 687.83 | 0.89 | 16.77 | 0.89 |
| Sri Veeramakaliamman Singapore | 121.02 | 157.39 | 2.19 | 16.87 (16.92) | 83.50 | 166.68 | 2.13 | 9.30 | 0.71 |
| Yueh Hai Ching Temple Singapore | 37.46 | 52.59 | 0.65 | 0.65 (1.16) | 25.60 | 51.35 | 0.65 | 0.73 | 0.65 |
| Average | 80.66 | 85.61 | 1.92 | 3.52 (7.07) | 80.69 | 436.51 | 0.85 | 4.86 | 0.66 |
| Rotation error (deg) | Alcatraz Courtyard | 10.102 | 13.201 | 2.607 | 0.035 | 6.093 | 10.946 | 0.038 | 0.030 | 0.043 |
| Alcatraz Water Tower | 10.637 | 11.053 | 0.499 | 0.764 | 11.501 | 10.641 | 0.699 | 19.351 | 0.228 |
| Drinking Fountain Somewhere in Zurich | 15.846 | 16.014 | 0.003 | 0.001 | 15.415 | 15.704 | 0.001 | 22.759 | 0.007 |
| Nijo Castle Gate | 16.751 | 10.546 | 0.062 | 0.062 | 17.347 | 20.032 | 0.038 | 0.036 | 0.064 |
| Porta San Donato Bologna | 23.839 | 24.120 | 0.095 | 0.094 | 18.411 | 25.004 | 0.094 | 0.094 | 0.099 |
| Round Church Cambridge | 18.906 | 14.473 | 0.029 | 0.026 | 10.295 | 18.685 | 0.029 | 1.086 | 0.035 |
| Smolny Cathedral St Petersburg | 19.387 | 17.971 | 0.023 | 0.022 | 11.662 | 14.380 | 0.023 | 0.019 | 0.029 |
| Some Cathedral in Barcelona | 27.270 | 30.471 | 10.009 | 20.050 | 27.908 | 29.119 | 0.020 | 47.892 | 0.025 |
| Sri Veeramakaliamman Singapore | 28.275 | 36.903 | 0.549 | 4.871 | 23.702 | 36.176 | 0.457 | 2.759 | 0.169 |
| Yueh Hai Ching Temple Singapore | 15.733 | 22.706 | 0.038 | 0.038 | 9.515 | 21.561 | 0.038 | 0.038 | 0.043 |
| Average | 18.675 | 19.746 | 1.391 | 2.596 | 15.185 | 20.225 | 0.144 | 9.406 | 0.074 |
| Translation error (m) | Alcatraz Courtyard | 4.82 | 4.93 | 1.09 | 0.01 | 2.73 | 5.74 | 0.01 | 0.01 | 0.01 |
| Alcatraz Water Tower | 8.66 | 7.35 | 0.31 | 0.44 | 7.53 | 7.77 | 0.41 | 9.05 | 0.12 |
| Drinking Fountain Somewhere in Zurich | 4.44 | 4.44 | 0.00 | 0.00 | 4.45 | 4.46 | 0.00 | 1.38 | 0.00 |
| Nijo Castle Gate | 5.07 | 3.08 | 0.01 | 0.01 | 5.67 | 6.95 | 0.01 | 0.01 | 0.01 |
| Porta San Donato Bologna | 9.50 | 10.72 | 0.05 | 0.05 | 4.80 | 10.48 | 0.05 | 0.05 | 0.05 |
| Round Church Cambridge | 8.94 | 7.19 | 0.01 | 0.01 | 5.28 | 9.00 | 0.01 | 0.56 | 0.01 |
| Smolny Cathedral St Petersburg | 2.70 | 2.43 | 0.01 | 0.01 | 2.33 | 2.52 | 0.01 | 0.01 | 0.01 |
| Some Cathedral in Barcelona | 12.64 | 12.69 | 3.22 | 6.38 | 12.32 | 12.66 | 0.01 | 11.93 | 0.01 |
| Sri Veeramakaliamman Singapore | 4.94 | 4.90 | 0.16 | 1.32 | 4.93 | 4.90 | 0.14 | 0.77 | 0.04 |
| Yueh Hai Ching Temple Singapore | 4.12 | 4.27 | 0.01 | 0.01 | 2.44 | 4.28 | 0.01 | 0.01 | 0.01 |
| Average | 6.58 | 6.20 | 0.49 | 0.82 | 5.25 | 6.88 | 0.07 | 2.38 | 0.03 |
+
+Table 1. Results of Euclidean reconstruction of novel test scenes, with and without data augmentation. The results of DPESFM [31] have been acquired by us training the model, along with the results reported by [31] in parentheses, if available. The result of Colmap, as reported by [31], is also added for reference.
+
+ | Without data augmentation | With data augmentation | Colmap |
| Inference | Inference + BA | Inference | Inference + BA |
| Ours | DPESFM | Ours | DPESFM | Ours | DPESFM | Ours | DPESFM |
| Reprojection error (px) | Alcatraz Courtyard | 68.41 | 68.50 | 1.70 | 0.81 (0.82) | 36.01 | 92.37 | 0.81 | 0.92 | 0.81 |
| Alcatraz Water Tower | 36.30 | 50.47 | 1.05 | 1.13 (0.55) | 87.67 | 2831.94 | 0.88 | 10.16 | 0.55 |
| Drinking Fountain Somewhere in Zurich | 36.02 | 45.87 | 0.55 | 0.31 (7.21) | 219.75 | 234.90 | 0.31 | 6.73 | 0.31 |
| Nijo Castle Gate | 65.71 | 64.53 | 0.73 | 0.73 (5.81) | 61.41 | 68.19 | 0.88 | 0.89 | 0.73 |
| Porta San Donato Bologna | 74.32 | 94.26 | 0.74 | 0.74 (1.10) | 52.15 | 84.46 | 0.76 | 0.75 | 0.75 |
| Round Church Cambridge | 61.18 | 55.51 | 0.39 | 0.39 (0.50) | 29.80 | 59.54 | 0.39 | 1.49 | 0.39 |
| Smolny Cathedral St Petersburg | 156.16 | 120.85 | 0.81 | 0.81 (15.15) | 85.38 | 87.81 | 0.81 | 0.81 | 0.81 |
| Some Cathedral in Barcelona | 150.01 | 146.10 | 10.42 | 12.71 (21.46) | 125.68 | 687.83 | 0.89 | 16.77 | 0.89 |
| Sri Veeramakaliamman Singapore | 121.02 | 157.39 | 2.19 | 16.87 (16.92) | 83.50 | 166.68 | 2.13 | 9.30 | 0.71 |
| Yueh Hai Ching Temple Singapore | 37.46 | 52.59 | 0.65 | 0.65 (1.16) | 25.60 | 51.35 | 0.65 | 0.73 | 0.65 |
| Average | 80.66 | 85.61 | 1.92 | 3.52 (7.07) | 80.69 | 436.51 | 0.85 | 4.86 | 0.66 |
| Rotation error (deg) | Alcatraz Courtyard | 10.102 | 13.201 | 2.607 | 0.035 | 6.093 | 10.946 | 0.038 | 0.030 | 0.043 |
| Alcatraz Water Tower | 10.637 | 11.053 | 0.499 | 0.764 | 11.501 | 10.641 | 0.699 | 19.351 | 0.228 |
| Drinking Fountain Somewhere in Zurich | 15.846 | 16.014 | 0.003 | 0.001 | 15.415 | 15.704 | 0.001 | 22.759 | 0.007 |
| Nijo Castle Gate | 16.751 | 10.546 | 0.062 | 0.062 | 17.347 | 20.032 | 0.038 | 0.036 | 0.064 |
| Porta San Donato Bologna | 23.839 | 24.120 | 0.095 | 0.094 | 18.411 | 25.004 | 0.094 | 0.094 | 0.099 |
| Round Church Cambridge | 18.906 | 14.473 | 0.029 | 0.026 | 10.295 | 18.685 | 0.029 | 1.086 | 0.035 |
| Smolny Cathedral St Petersburg | 19.387 | 17.971 | 0.023 | 0.022 | 11.662 | 14.380 | 0.023 | 0.019 | 0.029 |
| Some Cathedral in Barcelona | 27.270 | 30.471 | 10.009 | 20.050 | 27.908 | 29.119 | 0.020 | 47.892 | 0.025 |
| Sri Veeramakaliamman Singapore | 28.275 | 36.903 | 0.549 | 4.871 | 23.702 | 36.176 | 0.457 | 2.759 | 0.169 |
| Yueh Hai Ching Temple Singapore | 15.733 | 22.706 | 0.038 | 0.038 | 9.515 | 21.561 | 0.038 | 0.038 | 0.043 |
| Average | 18.675 | 19.746 | 1.391 | 2.596 | 15.185 | 20.225 | 0.144 | 9.406 | 0.074 |
| Translation error (m) | Alcatraz Courtyard | 4.82 | 4.93 | 1.09 | 0.01 | 2.73 | 5.74 | 0.01 | 0.01 | 0.01 |
| Alcatraz Water Tower | 8.66 | 7.35 | 0.31 | 0.44 | 7.53 | 7.77 | 0.41 | 9.05 | 0.12 |
| Drinking Fountain Somewhere in Zurich | 4.44 | 4.44 | 0.00 | 0.00 | 4.45 | 4.46 | 0.00 | 1.38 | 0.00 |
| Nijo Castle Gate | 5.07 | 3.08 | 0.01 | 0.01 | 5.67 | 6.95 | 0.01 | 0.01 | 0.01 |
| Porta San Donato Bologna | 9.50 | 10.72 | 0.05 | 0.05 | 4.80 | 10.48 | 0.05 | 0.05 | 0.05 |
| Round Church Cambridge | 8.94 | 7.19 | 0.01 | 0.01 | 5.28 | 9.00 | 0.01 | 0.56 | 0.01 |
| Smolny Cathedral St Petersburg | 2.70 | 2.43 | 0.01 | 0.01 | 2.33 | 2.52 | 0.01 | 0.01 | 0.01 |
| Some Cathedral in Barcelona | 12.64 | 12.69 | 3.22 | 6.38 | 12.32 | 12.66 | 0.01 | 11.93 | 0.01 |
| Sri Veeramakaliamman Singapore | 4.94 | 4.90 | 0.16 | 1.32 | 4.93 | 4.90 | 0.14 | 0.77 | 0.04 |
| Yueh Hai Ching Temple Singapore | 4.12 | 4.27 | 0.01 | 0.01 | 2.44 | 4.28 | 0.01 | 0.01 | 0.01 |
| Average | 6.58 | 6.20 | 0.49 | 0.82 | 5.25 | 6.88 | 0.07 | 2.38 | 0.03 |
",
+ "image_path": "b299fb1ecc507bbe1aeeb195be0886ffeef2f0e874c84281e49386e8679c0227.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 52,
+ 461,
+ 288,
+ 569
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 46,
+ 410,
+ 546,
+ 444
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 410,
+ 546,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 410,
+ 546,
+ 444
+ ],
+ "type": "text",
+ "content": "Table 1. Results of Euclidean reconstruction of novel test scenes, with and without data augmentation. The results of DPESFM [31] have been acquired by us training the model, along with the results reported by [31] in parentheses, if available. The result of Colmap, as reported by [31], is also added for reference."
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 52,
+ 461,
+ 288,
+ 569
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 52,
+ 461,
+ 288,
+ 569
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 52,
+ 461,
+ 288,
+ 569
+ ],
+ "type": "table",
+ "html": "| method | Pub. | Basic | MSE |
| RAPID [41] | ACMMM2014 | incorporate heterogeneous | 0.0200 |
| PAM[50] | ICCV2017 | residual-based, active learning | 0.0200 |
| ALamp[42] | CVPR2017 | layout-aware, multi-patch | 0.0190 |
| NIMA [63] | TIP2018 | predict distribution | 0.0210 |
| MPada [59] | ACMM2018 | attention, multi-patch | 0.0220 |
| MLSP [30] | CVPR2019 | staged training,multi-level features | 0.0190 |
| UIAA [73] | TIP2019 | unified probabilistic formulation | 0.0210 |
| HGCN [56] | CVPR2021 | graph convolution networks | 0.0200 |
| TANet [27] | IJCAI2022 | attention, adaptive features | 0.0161 |
| SAAN [72] | CVPR2023 | AdaIN,self-supervised Pretraining | 0.0185 |
| TDT(ours) | - | Triangular Distribution | 0.0172 |
",
+ "bbox": [
+ 504,
+ 113,
+ 890,
+ 250
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "weight pooling version. With additional convolutional layers, we obtain $\\mathcal{X}_g\\in \\mathbb{R}^{N\\times C\\times 8\\times 8}$ . When comparing with the $\\mathrm{FC^4}$ model, we randomly select 128 samples from the training set as the prior set. However, when comparing with the method proposed by Tang et al. [64] that utilizes the extended external sRGB datasets, we choose same images for scene classification as the prior set, but only 1280 samples.",
+ "bbox": [
+ 496,
+ 267,
+ 890,
+ 373
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Following previous AWB works[18, 31, 60], 3-fold cross-validation is used. Standard metrics (mean, median, tri-mean of all the errors, the mean of the lowest $25\\%$ , and the mean of the highest $25\\%$ of errors, the 95th percentile error) are reported in terms of angular error in degrees.",
+ "bbox": [
+ 496,
+ 373,
+ 890,
+ 450
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "4.2. Results and Analysis",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 462,
+ 694,
+ 478
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Facial Age Estimation",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 486,
+ 660,
+ 501
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "The quantitative comparison of a list of top-performing methods on the MegaAge-Asian Dataset is shown in Table 2. From the table, we can find that even stocked with pre-training on large-scale datasets, methods like [39, 78] are inferior to ours. Compared with SSR-Net, our improvement is over $8\\%$ . PVP [78] gets higher CA(3) and CA(5) indexes, due to the massive pretraining on ImageNet[11], IMDB-WIKI[53], and AFAD[45]. DAA[8], published in 2023, shares the same setting with our proposed method, while TDT obtains better scores in all CA indexes.",
+ "bbox": [
+ 496,
+ 502,
+ 890,
+ 652
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "To better prove the effectiveness of TDT, we also profile TDT on FG-Net [47], reported in Table 3. Using MAE as the metric, TDT leads the board with the smallest MAE, surpassing the same-setting competitor DAA.",
+ "bbox": [
+ 496,
+ 654,
+ 890,
+ 715
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Aesthetic Assessment.",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 717,
+ 656,
+ 729
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "In Table 4, we list the MSE scores for the selected and close-related aesthetic works. In an overall manner, TDT ranks second with an MSE score of 0.0172. This improvement simply originates from the deployment of the TDT plugin over SAAN [72], whose MSE is 0.0185, again showing the advantage of the label discriminative ability of TDT.",
+ "bbox": [
+ 496,
+ 732,
+ 890,
+ 823
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Illumination Estimation.",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 824,
+ 676,
+ 837
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "In Table 5, we report the statistics of the predicted angular error for a set of AWB algorithms on the Color Checker Dataset. On metrics like mean, and median values, the TDT based on $\\mathrm{FC^4}$ outperforms other state-of-the-art methods,",
+ "bbox": [
+ 496,
+ 839,
+ 890,
+ 898
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "11024",
+ "bbox": [
+ 478,
+ 925,
+ 519,
+ 936
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/947f0f3dd8e31845b569040e250ac333b31391bfb8da01b8c354eb2f24476b48.jpg",
+ "table_caption": [
+ "Table 5. Results on AWB Dataset Color Checker Dataset."
+ ],
+ "table_footnote": [],
+ "table_body": "| method | Pub. | Basic | MSE |
| RAPID [41] | ACMMM2014 | incorporate heterogeneous | 0.0200 |
| PAM[50] | ICCV2017 | residual-based, active learning | 0.0200 |
| ALamp[42] | CVPR2017 | layout-aware, multi-patch | 0.0190 |
| NIMA [63] | TIP2018 | predict distribution | 0.0210 |
| MPada [59] | ACMM2018 | attention, multi-patch | 0.0220 |
| MLSP [30] | CVPR2019 | staged training,multi-level features | 0.0190 |
| UIAA [73] | TIP2019 | unified probabilistic formulation | 0.0210 |
| HGCN [56] | CVPR2021 | graph convolution networks | 0.0200 |
| TANet [27] | IJCAI2022 | attention, adaptive features | 0.0161 |
| SAAN [72] | CVPR2023 | AdaIN,self-supervised Pretraining | 0.0185 |
| TDT(ours) | - | Triangular Distribution | 0.0172 |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.268,
+ 0.892,
+ 0.374
+ ],
+ "angle": 0,
+ "content": "weight pooling version. With additional convolutional layers, we obtain \\(\\mathcal{X}_g\\in \\mathbb{R}^{N\\times C\\times 8\\times 8}\\). When comparing with the \\(\\mathrm{FC^4}\\) model, we randomly select 128 samples from the training set as the prior set. However, when comparing with the method proposed by Tang et al. [64] that utilizes the extended external sRGB datasets, we choose same images for scene classification as the prior set, but only 1280 samples."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.374,
+ 0.892,
+ 0.451
+ ],
+ "angle": 0,
+ "content": "Following previous AWB works[18, 31, 60], 3-fold cross-validation is used. Standard metrics (mean, median, tri-mean of all the errors, the mean of the lowest \\(25\\%\\), and the mean of the highest \\(25\\%\\) of errors, the 95th percentile error) are reported in terms of angular error in degrees."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.463,
+ 0.695,
+ 0.479
+ ],
+ "angle": 0,
+ "content": "4.2. Results and Analysis"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.487,
+ 0.661,
+ 0.502
+ ],
+ "angle": 0,
+ "content": "Facial Age Estimation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.503,
+ 0.892,
+ 0.653
+ ],
+ "angle": 0,
+ "content": "The quantitative comparison of a list of top-performing methods on the MegaAge-Asian Dataset is shown in Table 2. From the table, we can find that even stocked with pre-training on large-scale datasets, methods like [39, 78] are inferior to ours. Compared with SSR-Net, our improvement is over \\(8\\%\\). PVP [78] gets higher CA(3) and CA(5) indexes, due to the massive pretraining on ImageNet[11], IMDB-WIKI[53], and AFAD[45]. DAA[8], published in 2023, shares the same setting with our proposed method, while TDT obtains better scores in all CA indexes."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.655,
+ 0.892,
+ 0.716
+ ],
+ "angle": 0,
+ "content": "To better prove the effectiveness of TDT, we also profile TDT on FG-Net [47], reported in Table 3. Using MAE as the metric, TDT leads the board with the smallest MAE, surpassing the same-setting competitor DAA."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.718,
+ 0.658,
+ 0.73
+ ],
+ "angle": 0,
+ "content": "Aesthetic Assessment."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.733,
+ 0.892,
+ 0.824
+ ],
+ "angle": 0,
+ "content": "In Table 4, we list the MSE scores for the selected and close-related aesthetic works. In an overall manner, TDT ranks second with an MSE score of 0.0172. This improvement simply originates from the deployment of the TDT plugin over SAAN [72], whose MSE is 0.0185, again showing the advantage of the label discriminative ability of TDT."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.825,
+ 0.677,
+ 0.838
+ ],
+ "angle": 0,
+ "content": "Illumination Estimation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.84,
+ 0.892,
+ 0.9
+ ],
+ "angle": 0,
+ "content": "In Table 5, we report the statistics of the predicted angular error for a set of AWB algorithms on the Color Checker Dataset. On metrics like mean, and median values, the TDT based on \\(\\mathrm{FC^4}\\) outperforms other state-of-the-art methods,"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.48,
+ 0.926,
+ 0.52,
+ 0.938
+ ],
+ "angle": 0,
+ "content": "11024"
+ }
+ ],
+ [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.102,
+ 0.09,
+ 0.444,
+ 0.103
+ ],
+ "angle": 0,
+ "content": "Table 5. Results on AWB Dataset Color Checker Dataset."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.082,
+ 0.115,
+ 0.468,
+ 0.233
+ ],
+ "angle": 0,
+ "content": "| method | Pub. | Basic | MSE |
| RAPID [41] | ACMMM2014 | incorporate heterogeneous | 0.0200 |
| PAM[50] | ICCV2017 | residual-based, active learning | 0.0200 |
| ALamp[42] | CVPR2017 | layout-aware, multi-patch | 0.0190 |
| NIMA [63] | TIP2018 | predict distribution | 0.0210 |
| MPada [59] | ACMM2018 | attention, multi-patch | 0.0220 |
| MLSP [30] | CVPR2019 | staged training,multi-level features | 0.0190 |
| UIAA [73] | TIP2019 | unified probabilistic formulation | 0.0210 |
| HGCN [56] | CVPR2021 | graph convolution networks | 0.0200 |
| TANet [27] | IJCAI2022 | attention, adaptive features | 0.0161 |
| SAAN [72] | CVPR2023 | AdaIN,self-supervised Pretraining | 0.0185 |
| TDT(ours) | - | Triangular Distribution | 0.0172 |
+
+weight pooling version. With additional convolutional layers, we obtain $\mathcal{X}_g\in \mathbb{R}^{N\times C\times 8\times 8}$ . When comparing with the $\mathrm{FC^4}$ model, we randomly select 128 samples from the training set as the prior set. However, when comparing with the method proposed by Tang et al. [64] that utilizes the extended external sRGB datasets, we choose same images for scene classification as the prior set, but only 1280 samples.
+
+Following previous AWB works[18, 31, 60], 3-fold cross-validation is used. Standard metrics (mean, median, tri-mean of all the errors, the mean of the lowest $25\%$ , and the mean of the highest $25\%$ of errors, the 95th percentile error) are reported in terms of angular error in degrees.
+
+# 4.2. Results and Analysis
+
+# Facial Age Estimation
+
+The quantitative comparison of a list of top-performing methods on the MegaAge-Asian Dataset is shown in Table 2. From the table, we can find that even stocked with pre-training on large-scale datasets, methods like [39, 78] are inferior to ours. Compared with SSR-Net, our improvement is over $8\%$ . PVP [78] gets higher CA(3) and CA(5) indexes, due to the massive pretraining on ImageNet[11], IMDB-WIKI[53], and AFAD[45]. DAA[8], published in 2023, shares the same setting with our proposed method, while TDT obtains better scores in all CA indexes.
+
+To better prove the effectiveness of TDT, we also profile TDT on FG-Net [47], reported in Table 3. Using MAE as the metric, TDT leads the board with the smallest MAE, surpassing the same-setting competitor DAA.
+
+# Aesthetic Assessment.
+
+In Table 4, we list the MSE scores for the selected and close-related aesthetic works. In an overall manner, TDT ranks second with an MSE score of 0.0172. This improvement simply originates from the deployment of the TDT plugin over SAAN [72], whose MSE is 0.0185, again showing the advantage of the label discriminative ability of TDT.
+
+# Illumination Estimation.
+
+In Table 5, we report the statistics of the predicted angular error for a set of AWB algorithms on the Color Checker Dataset. On metrics like mean, and median values, the TDT based on $\mathrm{FC^4}$ outperforms other state-of-the-art methods,
+
+Table 5. Results on AWB Dataset Color Checker Dataset.
+
+| method | Pub. | Basic | MSE |
| RAPID [41] | ACMMM2014 | incorporate heterogeneous | 0.0200 |
| PAM[50] | ICCV2017 | residual-based, active learning | 0.0200 |
| ALamp[42] | CVPR2017 | layout-aware, multi-patch | 0.0190 |
| NIMA [63] | TIP2018 | predict distribution | 0.0210 |
| MPada [59] | ACMM2018 | attention, multi-patch | 0.0220 |
| MLSP [30] | CVPR2019 | staged training,multi-level features | 0.0190 |
| UIAA [73] | TIP2019 | unified probabilistic formulation | 0.0210 |
| HGCN [56] | CVPR2021 | graph convolution networks | 0.0200 |
| TANet [27] | IJCAI2022 | attention, adaptive features | 0.0161 |
| SAAN [72] | CVPR2023 | AdaIN,self-supervised Pretraining | 0.0185 |
| TDT(ours) | - | Triangular Distribution | 0.0172 |
",
+ "image_path": "13d482d354dfabaa7eebe2528e084fe73229e4f203605de741f244fa7ffd7cc8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 304,
+ 212,
+ 545,
+ 296
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 212,
+ 545,
+ 296
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 212,
+ 545,
+ 296
+ ],
+ "type": "text",
+ "content": "weight pooling version. With additional convolutional layers, we obtain "
+ },
+ {
+ "bbox": [
+ 304,
+ 212,
+ 545,
+ 296
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{X}_g\\in \\mathbb{R}^{N\\times C\\times 8\\times 8}"
+ },
+ {
+ "bbox": [
+ 304,
+ 212,
+ 545,
+ 296
+ ],
+ "type": "text",
+ "content": ". When comparing with the "
+ },
+ {
+ "bbox": [
+ 304,
+ 212,
+ 545,
+ 296
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{FC^4}"
+ },
+ {
+ "bbox": [
+ 304,
+ 212,
+ 545,
+ 296
+ ],
+ "type": "text",
+ "content": " model, we randomly select 128 samples from the training set as the prior set. However, when comparing with the method proposed by Tang et al. [64] that utilizes the extended external sRGB datasets, we choose same images for scene classification as the prior set, but only 1280 samples."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 304,
+ 296,
+ 545,
+ 357
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 296,
+ 545,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 296,
+ 545,
+ 357
+ ],
+ "type": "text",
+ "content": "Following previous AWB works[18, 31, 60], 3-fold cross-validation is used. Standard metrics (mean, median, tri-mean of all the errors, the mean of the lowest "
+ },
+ {
+ "bbox": [
+ 304,
+ 296,
+ 545,
+ 357
+ ],
+ "type": "inline_equation",
+ "content": "25\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 296,
+ 545,
+ 357
+ ],
+ "type": "text",
+ "content": ", and the mean of the highest "
+ },
+ {
+ "bbox": [
+ 304,
+ 296,
+ 545,
+ 357
+ ],
+ "type": "inline_equation",
+ "content": "25\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 296,
+ 545,
+ 357
+ ],
+ "type": "text",
+ "content": " of errors, the 95th percentile error) are reported in terms of angular error in degrees."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 306,
+ 366,
+ 425,
+ 379
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 366,
+ 425,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 366,
+ 425,
+ 379
+ ],
+ "type": "text",
+ "content": "4.2. Results and Analysis"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 306,
+ 385,
+ 404,
+ 397
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 385,
+ 404,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 385,
+ 404,
+ 397
+ ],
+ "type": "text",
+ "content": "Facial Age Estimation"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 304,
+ 398,
+ 545,
+ 517
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 398,
+ 545,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 398,
+ 545,
+ 517
+ ],
+ "type": "text",
+ "content": "The quantitative comparison of a list of top-performing methods on the MegaAge-Asian Dataset is shown in Table 2. From the table, we can find that even stocked with pre-training on large-scale datasets, methods like [39, 78] are inferior to ours. Compared with SSR-Net, our improvement is over "
+ },
+ {
+ "bbox": [
+ 304,
+ 398,
+ 545,
+ 517
+ ],
+ "type": "inline_equation",
+ "content": "8\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 398,
+ 545,
+ 517
+ ],
+ "type": "text",
+ "content": ". PVP [78] gets higher CA(3) and CA(5) indexes, due to the massive pretraining on ImageNet[11], IMDB-WIKI[53], and AFAD[45]. DAA[8], published in 2023, shares the same setting with our proposed method, while TDT obtains better scores in all CA indexes."
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 304,
+ 518,
+ 545,
+ 567
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 518,
+ 545,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 518,
+ 545,
+ 567
+ ],
+ "type": "text",
+ "content": "To better prove the effectiveness of TDT, we also profile TDT on FG-Net [47], reported in Table 3. Using MAE as the metric, TDT leads the board with the smallest MAE, surpassing the same-setting competitor DAA."
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 306,
+ 568,
+ 402,
+ 578
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 568,
+ 402,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 568,
+ 402,
+ 578
+ ],
+ "type": "text",
+ "content": "Aesthetic Assessment."
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 304,
+ 580,
+ 545,
+ 652
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 580,
+ 545,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 580,
+ 545,
+ 652
+ ],
+ "type": "text",
+ "content": "In Table 4, we list the MSE scores for the selected and close-related aesthetic works. In an overall manner, TDT ranks second with an MSE score of 0.0172. This improvement simply originates from the deployment of the TDT plugin over SAAN [72], whose MSE is 0.0185, again showing the advantage of the label discriminative ability of TDT."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 306,
+ 653,
+ 414,
+ 663
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 653,
+ 414,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 653,
+ 414,
+ 663
+ ],
+ "type": "text",
+ "content": "Illumination Estimation."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 304,
+ 665,
+ 545,
+ 712
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 665,
+ 545,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 665,
+ 545,
+ 712
+ ],
+ "type": "text",
+ "content": "In Table 5, we report the statistics of the predicted angular error for a set of AWB algorithms on the Color Checker Dataset. On metrics like mean, and median values, the TDT based on "
+ },
+ {
+ "bbox": [
+ 304,
+ 665,
+ 545,
+ 712
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{FC^4}"
+ },
+ {
+ "bbox": [
+ 304,
+ 665,
+ 545,
+ 712
+ ],
+ "type": "text",
+ "content": " outperforms other state-of-the-art methods,"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 293,
+ 733,
+ 318,
+ 742
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 293,
+ 733,
+ 318,
+ 742
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 293,
+ 733,
+ 318,
+ 742
+ ],
+ "type": "text",
+ "content": "11024"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 50,
+ 91,
+ 286,
+ 184
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 62,
+ 71,
+ 271,
+ 81
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 62,
+ 71,
+ 271,
+ 81
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 62,
+ 71,
+ 271,
+ 81
+ ],
+ "type": "text",
+ "content": "Table 5. Results on AWB Dataset Color Checker Dataset."
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 50,
+ 91,
+ 286,
+ 184
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 91,
+ 286,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 91,
+ 286,
+ 184
+ ],
+ "type": "table",
+ "html": "| Templates | In context examples |
| c->caption | revolver ->Multiple antique revolvers lie on a wooden table, gleaming under soft, ambient light. closetet ->The compact closet, brimming with clothes and shoes, exudes a feeling of organization. zebra ->A zebra is gallantly trotting across the vast, sunlit plains of the African savannah, creating a captivating black and white spectacle. bus station ->The bustling bus station thrums with restless energy, as travelers navigate through the crowded space, awaiting their journeys amid the echoes of departing buses. |
| c,bg->caption | tiger, forest ->Two tigers are running together in the forest. lighter, motorhome ->In the cozy, cluttered environment of a well-traveled motorhome, a sleek silver lighter holds dominion on the rustic wooden table. sunset, lake ->Golden sunset hues reflect on a calm lake, silhouetting a lone canoeist against a backdrop of fiery clouds. |
| c,rel->caption | kit fox, in front of ->A group of small, fluffy, golden kit foxes is playfully gathered in front of a lush, green, towering forest backdrop. cabbage, besides ->A vibrant image portrays a lush, green cabbage, glistening with dewdrops, nestled besides a rustic, wooden crate full of freshly harvested vegetables. |
",
+ "bbox": [
+ 89,
+ 88,
+ 880,
+ 323
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Table 1. We show examples for the three synthesis templates. Such examples are used as demonstrations for Llama-2 to perform the in-context learning task. We have 176 such examples in total. Most of them are generated by prompting GPT-4 [58], while a handful of others are human generated (in a 10M scale pilot study of synthetic captions, we do not notice significant differences between including or excluding human generated examples.)",
+ "bbox": [
+ 75,
+ 330,
+ 893,
+ 388
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "image",
+ "img_path": "images/8a42f1841f22903e4b63eb70e60d711d71f5a05e01cac9bfe87c99dfd12c5267.jpg",
+ "image_caption": [
+ "Figure 3. In-context caption generation using Llama-2 [83]. We randomly sample three in-context examples for each inference run."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 81,
+ 407,
+ 467,
+ 523
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "of, our objective is to prompt the LLM to create captions such as a cute yellow cat is enjoying the fish in front of the sofa. To add variety, we have a selection of 10 different positional relationship words that we randomly choose from.",
+ "bbox": [
+ 89,
+ 594,
+ 468,
+ 670
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "For each of the three templates, we have prepared multiple demonstration examples that serve as instructions for the LLM to complete the caption synthesis task. Table 1 shows a couple of examples for each template. In total, we have 106 examples for $c->prompt$ , 50 examples for $c, bg->prompt$ , and 20 examples for $c, rel->prompt$ . Such examples are mostly collected by prompting GPT-4, with a handful from human. In a pilot study, we do not observe difference between including or excluding human generated examples.",
+ "bbox": [
+ 75,
+ 683,
+ 470,
+ 820
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "In the stage of generating captions in-context, we select a concept and one of the three templates. Next, we randomly pick three examples from the chosen template and frame the caption generation as a text completion task. This process is illustrated in Figure 3.",
+ "bbox": [
+ 75,
+ 825,
+ 470,
+ 902
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "3.2. Synthesizing Images",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 407,
+ 692,
+ 424
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "For each text caption, we generate a variety of images by initiating the reverse diffusion process with different random noise. The Classifier-Free Guidance (CFG) scale is a crucial factor in this process. A higher CFG scale enhances the quality of the samples and the alignment between text and image, whereas a lower scale results in more diverse samples and better adherence to the original conditional distribution of images based on the given text. Following the approach used in StableRep [81], we opt for a lower CFG scale, specifically 2.5, and produce 4 images for each caption. Examples of these images can be seen in Figure 4.",
+ "bbox": [
+ 496,
+ 436,
+ 893,
+ 604
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "3.3. Representation Learning",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 635,
+ 728,
+ 652
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Our representation learning method is built upon StableRep [81]. The key component of our approach is the multi-positive contrastive learning loss [43] which works by aligning (in the embedding space) images generated from the same caption. We additionally combine multiple techniques from other self-supervised learning methods, including a patch-level masked image modeling objective. We briefly review StableRep and elaborate on the added modules.",
+ "bbox": [
+ 496,
+ 666,
+ 893,
+ 787
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "StableRep [81] minimizes the cross-entropy loss between a ground-truth assignment distribution and a contrastive assignment distribution. Consider an encoded anchor sample $\\pmb{a}$ and a set of encoded candidates $\\{\\pmb{b}_1, \\pmb{b}_2, \\dots, \\pmb{b}_K\\}$ . The contrastive assignment distribution $\\mathbf{q}$ describes how likely the model predicts $\\pmb{a}$ and each $\\pmb{b}$ to be generated from the same caption, and the ground-truth distribution is the actual match",
+ "bbox": [
+ 496,
+ 794,
+ 895,
+ 902
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "15890",
+ "bbox": [
+ 480,
+ 944,
+ 519,
+ 957
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "image",
+ "img_path": "images/dc99af57ce94112007c0043dbe70f56a359cc6068754a28f472dc85053281589.jpg",
+ "image_caption": [
+ "A plate of paella, a mixed rice dish with chicken, beans, and seafood"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 99,
+ 88,
+ 447,
+ 157
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/9c549d9d6852c2bed63f35884b2498ecbe23a23b1be85f1137677ab6c2fb8d44.jpg",
+ "image_caption": [
+ "An industrial power plant with its smokestacks belching black smoke."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 99,
+ 174,
+ 447,
+ 243
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/7183cd8ec8c687dba982a5f8c99be5008b3add07a27b232ead4282a53c4e14f9.jpg",
+ "image_caption": [
+ "A fluffy, black and white junco bird perches on a snow-covered fence, overlooking a dark forest.",
+ "Figure 4. Random examples of synthetic captions and images generated in our SynCLR pipeline. Each caption comes with 4 images."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 99,
+ 253,
+ 447,
+ 321
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "between $\\pmb{a}$ and $\\pmb{b}$ ( $\\pmb{a}$ is allowed to match multiple $\\pmb{b}$ ):",
+ "bbox": [
+ 76,
+ 404,
+ 421,
+ 420
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {q} _ {i} = \\frac {\\exp \\left(\\boldsymbol {a} \\cdot \\boldsymbol {b} _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\boldsymbol {a} \\cdot \\boldsymbol {b} _ {j} / \\tau\\right)} \\tag {1}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 181,
+ 426,
+ 468,
+ 465
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathbf {p} _ {i} = \\frac {\\mathbb {1} _ {\\operatorname {m a t c h} \\left(\\boldsymbol {a} , \\boldsymbol {b} _ {i}\\right)}}{\\sum_ {j = 1} ^ {K} \\mathbb {1} _ {\\operatorname {m a t c h} \\left(\\boldsymbol {a} , \\boldsymbol {b} _ {j}\\right)}} \\tag {2}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 200,
+ 467,
+ 468,
+ 503
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $\\tau \\in \\mathcal{R}_{+}$ is the scalar temperature, $\\pmb{a}$ and all $\\pmb{b}$ have been $\\ell_{2}$ normalized, and the indicator function $\\mathbb{1}_{\\mathrm{match}(\\cdot ,\\cdot)}$ indicates whether two samples are from the same caption. The contrastive loss for $\\pmb{a}$ is given as",
+ "bbox": [
+ 76,
+ 511,
+ 468,
+ 571
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} (\\boldsymbol {a}) = H (\\boldsymbol {p}, \\boldsymbol {q}) = - \\sum_ {i = 1} ^ {K} \\mathbf {p} _ {i} \\log \\mathbf {q} _ {i} \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 153,
+ 580,
+ 468,
+ 622
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "iBOT [98] is a masked image modeling objective, wherein a localized patch is masked, and the model is tasked with predicting the tokenized representation of said masked patch. It adapts the DINO [11] objective from the image level into the patch level. We follow [67] to replace the softmax-centering method with the iterative Sinkhorn-Knopp (SK) algorithm [19]. We run SK for 3 iterations to build the prediction target.",
+ "bbox": [
+ 75,
+ 628,
+ 468,
+ 750
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Exponential Moving Average (EMA) is firstly introduced into self-supervised learning by MoCo [38]. We use EMA to encode crops as $\\mathbf{b}$ and to produce the targets for iBOT loss. We update the EMA model as $\\theta_{ema} \\gets \\lambda \\theta_{ema} + (1 - \\lambda)\\theta$ , following a cosine schedule for $\\lambda$ from 0.994 to 1 during training [34, 59]. We find the EMA module not only increases the final performance, but also improves the training stability for long training schedules.",
+ "bbox": [
+ 75,
+ 750,
+ 470,
+ 869
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Multi-crop strategy is introduced by [10] as a smart way to improve computation efficiency, and is adopted in this paper.",
+ "bbox": [
+ 76,
+ 869,
+ 470,
+ 901
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "For these local crops, we only employ the contrastive loss, omitting the iBOT loss. Local crops are encoded only by the student network, and matched to global crops from the same caption encoded by the EMA model. Such reuse of global crops saves computation. For each image $x$ , where we generate a single global crop $x^g$ alongside $n$ local crops $x^l$ , the final loss can be expressed as follows:",
+ "bbox": [
+ 496,
+ 90,
+ 893,
+ 196
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} \\left(x ^ {g}\\right) + \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathcal {L} \\left(x _ {i} ^ {l}\\right) + \\mathcal {L} ^ {i B O T} \\left(x ^ {g}\\right) \\tag {4}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 573,
+ 204,
+ 892,
+ 243
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.4. Implementation",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 250,
+ 660,
+ 266
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Concept list. We concatenate class names from various datasets, including IN-1k [21], IN-21k (we keep the most frequent 13k classes), Aircraft [51], Cars [44], DTD [17], Flowers [55], Pets [60], Sun397 [88], Caltech-101 [30], Food101 [7], and Places-365 [96]. If the concept is a place (i.e. SUN397 and Places) or a texture (i.e. DTD), we only apply the $c \\rightarrow$ caption template. For fine-grained classes such as pets or flowers, we employ GPT-4 to generate a consolidated list of probable backgrounds, rather than producing distinct lists for each specific class. We favor more frequent sampling from IN-1k, Food101, Cars, Aircraft, and Flowers.",
+ "bbox": [
+ 496,
+ 273,
+ 893,
+ 438
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Batches. For each training batch, we sample 2048 captions (except when noted), and use all of the 4 images generated by each caption. We generate 1 global and 4 local crops for each image. As a result, each batch contains 8192 global crops, which is similar with prior work [13, 14, 34, 81].",
+ "bbox": [
+ 496,
+ 439,
+ 893,
+ 515
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Masking. For the iBOT loss, we randomly choose $50\\%$ images inside a batch to mask, and randomly mask $50\\%$ of the tokens in each chosen image. We use 65536 prototypes. While the target from the EMA model is ascertained using the SK algorithm, we apply softmax normalization to the output of the student model.",
+ "bbox": [
+ 496,
+ 513,
+ 893,
+ 604
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Projection heads. We follow the design in MoCo v3 [14] and DINO [11] for the contrastive and iBOT loss heads, respectively, ensuring consistency with established methods. Other hyper-parameters. We set the temperature in the contrastive loss to 0.08. For the temperature used in the iBOT loss, we linearly increase it from 0.04 to 0.07 over 4000 iterations, and keep it as 0.07 afterwards, as in DINO [11]. Additionally, the weight decay parameter is incrementally adjusted from 0.04 to 0.2, adhering to a cosine schedule.",
+ "bbox": [
+ 496,
+ 604,
+ 893,
+ 741
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4. Experiment",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 753,
+ 624,
+ 770
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "We first perform an ablation study to evaluate the efficacy of various designs and modules within our pipeline. Then we proceed to scale up the volume of synthetic data.",
+ "bbox": [
+ 496,
+ 777,
+ 892,
+ 824
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "4.1. Study different components",
+ "text_level": 1,
+ "bbox": [
+ 500,
+ 832,
+ 750,
+ 848
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "We analyze each component of SynCLR, and ablate their effectiveness in two measurements: (1) linear probing performance on IN-1k; (2) average accuracy of linear transfer on",
+ "bbox": [
+ 496,
+ 854,
+ 893,
+ 900
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "15891",
+ "bbox": [
+ 480,
+ 944,
+ 517,
+ 955
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "table",
+ "img_path": "images/6620bcbf13362b40399a4909011b5b3d32e0506477099ad68d39c46b1c083af4.jpg",
+ "table_caption": [],
+ "table_footnote": [],
+ "table_body": "| Templates | In context examples |
| c->caption | revolver ->Multiple antique revolvers lie on a wooden table, gleaming under soft, ambient light. closetet ->The compact closet, brimming with clothes and shoes, exudes a feeling of organization. zebra ->A zebra is gallantly trotting across the vast, sunlit plains of the African savannah, creating a captivating black and white spectacle. bus station ->The bustling bus station thrums with restless energy, as travelers navigate through the crowded space, awaiting their journeys amid the echoes of departing buses. |
| c,bg->caption | tiger, forest ->Two tigers are running together in the forest. lighter, motorhome ->In the cozy, cluttered environment of a well-traveled motorhome, a sleek silver lighter holds dominion on the rustic wooden table. sunset, lake ->Golden sunset hues reflect on a calm lake, silhouetting a lone canoeist against a backdrop of fiery clouds. |
| c,rel->caption | kit fox, in front of ->A group of small, fluffy, golden kit foxes is playfully gathered in front of a lush, green, towering forest backdrop. cabbage, besides ->A vibrant image portrays a lush, green cabbage, glistening with dewdrops, nestled besides a rustic, wooden crate full of freshly harvested vegetables. |
"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.076,
+ 0.332,
+ 0.895,
+ 0.39
+ ],
+ "angle": 0,
+ "content": "Table 1. We show examples for the three synthesis templates. Such examples are used as demonstrations for Llama-2 to perform the in-context learning task. We have 176 such examples in total. Most of them are generated by prompting GPT-4 [58], while a handful of others are human generated (in a 10M scale pilot study of synthetic captions, we do not notice significant differences between including or excluding human generated examples.)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.082,
+ 0.408,
+ 0.468,
+ 0.525
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.077,
+ 0.535,
+ 0.472,
+ 0.565
+ ],
+ "angle": 0,
+ "content": "Figure 3. In-context caption generation using Llama-2 [83]. We randomly sample three in-context examples for each inference run."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.09,
+ 0.595,
+ 0.47,
+ 0.671
+ ],
+ "angle": 0,
+ "content": "of, our objective is to prompt the LLM to create captions such as a cute yellow cat is enjoying the fish in front of the sofa. To add variety, we have a selection of 10 different positional relationship words that we randomly choose from."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.684,
+ 0.472,
+ 0.821
+ ],
+ "angle": 0,
+ "content": "For each of the three templates, we have prepared multiple demonstration examples that serve as instructions for the LLM to complete the caption synthesis task. Table 1 shows a couple of examples for each template. In total, we have 106 examples for \\( c->prompt \\), 50 examples for \\( c, bg->prompt \\), and 20 examples for \\( c, rel->prompt \\). Such examples are mostly collected by prompting GPT-4, with a handful from human. In a pilot study, we do not observe difference between including or excluding human generated examples."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.826,
+ 0.471,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "In the stage of generating captions in-context, we select a concept and one of the three templates. Next, we randomly pick three examples from the chosen template and frame the caption generation as a text completion task. This process is illustrated in Figure 3."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.408,
+ 0.694,
+ 0.425
+ ],
+ "angle": 0,
+ "content": "3.2. Synthesizing Images"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.438,
+ 0.895,
+ 0.606
+ ],
+ "angle": 0,
+ "content": "For each text caption, we generate a variety of images by initiating the reverse diffusion process with different random noise. The Classifier-Free Guidance (CFG) scale is a crucial factor in this process. A higher CFG scale enhances the quality of the samples and the alignment between text and image, whereas a lower scale results in more diverse samples and better adherence to the original conditional distribution of images based on the given text. Following the approach used in StableRep [81], we opt for a lower CFG scale, specifically 2.5, and produce 4 images for each caption. Examples of these images can be seen in Figure 4."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.636,
+ 0.729,
+ 0.654
+ ],
+ "angle": 0,
+ "content": "3.3. Representation Learning"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.667,
+ 0.895,
+ 0.789
+ ],
+ "angle": 0,
+ "content": "Our representation learning method is built upon StableRep [81]. The key component of our approach is the multi-positive contrastive learning loss [43] which works by aligning (in the embedding space) images generated from the same caption. We additionally combine multiple techniques from other self-supervised learning methods, including a patch-level masked image modeling objective. We briefly review StableRep and elaborate on the added modules."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.795,
+ 0.897,
+ 0.903
+ ],
+ "angle": 0,
+ "content": "StableRep [81] minimizes the cross-entropy loss between a ground-truth assignment distribution and a contrastive assignment distribution. Consider an encoded anchor sample \\(\\pmb{a}\\) and a set of encoded candidates \\(\\{\\pmb{b}_1, \\pmb{b}_2, \\dots, \\pmb{b}_K\\}\\). The contrastive assignment distribution \\(\\mathbf{q}\\) describes how likely the model predicts \\(\\pmb{a}\\) and each \\(\\pmb{b}\\) to be generated from the same caption, and the ground-truth distribution is the actual match"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.52,
+ 0.958
+ ],
+ "angle": 0,
+ "content": "15890"
+ }
+ ],
+ [
+ {
+ "type": "image",
+ "bbox": [
+ 0.1,
+ 0.089,
+ 0.449,
+ 0.158
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.102,
+ 0.157,
+ 0.405,
+ 0.166
+ ],
+ "angle": 0,
+ "content": "A plate of paella, a mixed rice dish with chicken, beans, and seafood"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.1,
+ 0.175,
+ 0.449,
+ 0.244
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.102,
+ 0.244,
+ 0.408,
+ 0.252
+ ],
+ "angle": 0,
+ "content": "An industrial power plant with its smokestacks belching black smoke."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.1,
+ 0.254,
+ 0.449,
+ 0.323
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.102,
+ 0.323,
+ 0.414,
+ 0.34
+ ],
+ "angle": 0,
+ "content": "A fluffy, black and white junco bird perches on a snow-covered fence, overlooking a dark forest."
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.077,
+ 0.346,
+ 0.47,
+ 0.387
+ ],
+ "angle": 0,
+ "content": "Figure 4. Random examples of synthetic captions and images generated in our SynCLR pipeline. Each caption comes with 4 images."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.405,
+ 0.423,
+ 0.421
+ ],
+ "angle": 0,
+ "content": "between \\( \\pmb{a} \\) and \\( \\pmb{b} \\) (\\( \\pmb{a} \\) is allowed to match multiple \\( \\pmb{b} \\)):"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.182,
+ 0.428,
+ 0.47,
+ 0.467
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathbf {q} _ {i} = \\frac {\\exp \\left(\\boldsymbol {a} \\cdot \\boldsymbol {b} _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\boldsymbol {a} \\cdot \\boldsymbol {b} _ {j} / \\tau\\right)} \\tag {1}\n\\]"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.202,
+ 0.468,
+ 0.47,
+ 0.505
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathbf {p} _ {i} = \\frac {\\mathbb {1} _ {\\operatorname {m a t c h} \\left(\\boldsymbol {a} , \\boldsymbol {b} _ {i}\\right)}}{\\sum_ {j = 1} ^ {K} \\mathbb {1} _ {\\operatorname {m a t c h} \\left(\\boldsymbol {a} , \\boldsymbol {b} _ {j}\\right)}} \\tag {2}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.512,
+ 0.47,
+ 0.573
+ ],
+ "angle": 0,
+ "content": "where \\(\\tau \\in \\mathcal{R}_{+}\\) is the scalar temperature, \\(\\pmb{a}\\) and all \\(\\pmb{b}\\) have been \\(\\ell_{2}\\) normalized, and the indicator function \\(\\mathbb{1}_{\\mathrm{match}(\\cdot ,\\cdot)}\\) indicates whether two samples are from the same caption. The contrastive loss for \\(\\pmb{a}\\) is given as"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.154,
+ 0.581,
+ 0.47,
+ 0.623
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} (\\boldsymbol {a}) = H (\\boldsymbol {p}, \\boldsymbol {q}) = - \\sum_ {i = 1} ^ {K} \\mathbf {p} _ {i} \\log \\mathbf {q} _ {i} \\tag {3}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.629,
+ 0.47,
+ 0.75
+ ],
+ "angle": 0,
+ "content": "iBOT [98] is a masked image modeling objective, wherein a localized patch is masked, and the model is tasked with predicting the tokenized representation of said masked patch. It adapts the DINO [11] objective from the image level into the patch level. We follow [67] to replace the softmax-centering method with the iterative Sinkhorn-Knopp (SK) algorithm [19]. We run SK for 3 iterations to build the prediction target."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.75,
+ 0.471,
+ 0.871
+ ],
+ "angle": 0,
+ "content": "Exponential Moving Average (EMA) is firstly introduced into self-supervised learning by MoCo [38]. We use EMA to encode crops as \\( \\mathbf{b} \\) and to produce the targets for iBOT loss. We update the EMA model as \\( \\theta_{ema} \\gets \\lambda \\theta_{ema} + (1 - \\lambda)\\theta \\), following a cosine schedule for \\( \\lambda \\) from 0.994 to 1 during training [34, 59]. We find the EMA module not only increases the final performance, but also improves the training stability for long training schedules."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.077,
+ 0.871,
+ 0.472,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "Multi-crop strategy is introduced by [10] as a smart way to improve computation efficiency, and is adopted in this paper."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.092,
+ 0.894,
+ 0.198
+ ],
+ "angle": 0,
+ "content": "For these local crops, we only employ the contrastive loss, omitting the iBOT loss. Local crops are encoded only by the student network, and matched to global crops from the same caption encoded by the EMA model. Such reuse of global crops saves computation. For each image \\( x \\), where we generate a single global crop \\( x^g \\) alongside \\( n \\) local crops \\( x^l \\), the final loss can be expressed as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.574,
+ 0.205,
+ 0.893,
+ 0.244
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} \\left(x ^ {g}\\right) + \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathcal {L} \\left(x _ {i} ^ {l}\\right) + \\mathcal {L} ^ {i B O T} \\left(x ^ {g}\\right) \\tag {4}\n\\]"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.251,
+ 0.661,
+ 0.267
+ ],
+ "angle": 0,
+ "content": "3.4. Implementation"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.274,
+ 0.895,
+ 0.439
+ ],
+ "angle": 0,
+ "content": "Concept list. We concatenate class names from various datasets, including IN-1k [21], IN-21k (we keep the most frequent 13k classes), Aircraft [51], Cars [44], DTD [17], Flowers [55], Pets [60], Sun397 [88], Caltech-101 [30], Food101 [7], and Places-365 [96]. If the concept is a place (i.e. SUN397 and Places) or a texture (i.e. DTD), we only apply the \\( c \\rightarrow \\) caption template. For fine-grained classes such as pets or flowers, we employ GPT-4 to generate a consolidated list of probable backgrounds, rather than producing distinct lists for each specific class. We favor more frequent sampling from IN-1k, Food101, Cars, Aircraft, and Flowers."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.44,
+ 0.894,
+ 0.516
+ ],
+ "angle": 0,
+ "content": "Batches. For each training batch, we sample 2048 captions (except when noted), and use all of the 4 images generated by each caption. We generate 1 global and 4 local crops for each image. As a result, each batch contains 8192 global crops, which is similar with prior work [13, 14, 34, 81]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.515,
+ 0.894,
+ 0.605
+ ],
+ "angle": 0,
+ "content": "Masking. For the iBOT loss, we randomly choose \\(50\\%\\) images inside a batch to mask, and randomly mask \\(50\\%\\) of the tokens in each chosen image. We use 65536 prototypes. While the target from the EMA model is ascertained using the SK algorithm, we apply softmax normalization to the output of the student model."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.606,
+ 0.894,
+ 0.742
+ ],
+ "angle": 0,
+ "content": "Projection heads. We follow the design in MoCo v3 [14] and DINO [11] for the contrastive and iBOT loss heads, respectively, ensuring consistency with established methods. Other hyper-parameters. We set the temperature in the contrastive loss to 0.08. For the temperature used in the iBOT loss, we linearly increase it from 0.04 to 0.07 over 4000 iterations, and keep it as 0.07 afterwards, as in DINO [11]. Additionally, the weight decay parameter is incrementally adjusted from 0.04 to 0.2, adhering to a cosine schedule."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.754,
+ 0.625,
+ 0.771
+ ],
+ "angle": 0,
+ "content": "4. Experiment"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.779,
+ 0.893,
+ 0.825
+ ],
+ "angle": 0,
+ "content": "We first perform an ablation study to evaluate the efficacy of various designs and modules within our pipeline. Then we proceed to scale up the volume of synthetic data."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.5,
+ 0.833,
+ 0.75,
+ 0.849
+ ],
+ "angle": 0,
+ "content": "4.1. Study different components"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.498,
+ 0.856,
+ 0.894,
+ 0.901
+ ],
+ "angle": 0,
+ "content": "We analyze each component of SynCLR, and ablate their effectiveness in two measurements: (1) linear probing performance on IN-1k; (2) average accuracy of linear transfer on"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.481,
+ 0.945,
+ 0.518,
+ 0.957
+ ],
+ "angle": 0,
+ "content": "15891"
+ }
+ ],
+ [
+ {
+ "type": "table",
+ "bbox": [
+ 0.12,
+ 0.076,
+ 0.428,
+ 0.196
+ ],
+ "angle": 0,
+ "content": "| Templates | In context examples |
| c->caption | revolver ->Multiple antique revolvers lie on a wooden table, gleaming under soft, ambient light. closetet ->The compact closet, brimming with clothes and shoes, exudes a feeling of organization. zebra ->A zebra is gallantly trotting across the vast, sunlit plains of the African savannah, creating a captivating black and white spectacle. bus station ->The bustling bus station thrums with restless energy, as travelers navigate through the crowded space, awaiting their journeys amid the echoes of departing buses. |
| c,bg->caption | tiger, forest ->Two tigers are running together in the forest. lighter, motorhome ->In the cozy, cluttered environment of a well-traveled motorhome, a sleek silver lighter holds dominion on the rustic wooden table. sunset, lake ->Golden sunset hues reflect on a calm lake, silhouetting a lone canoeist against a backdrop of fiery clouds. |
| c,rel->caption | kit fox, in front of ->A group of small, fluffy, golden kit foxes is playfully gathered in front of a lush, green, towering forest backdrop. cabbage, besides ->A vibrant image portrays a lush, green cabbage, glistening with dewdrops, nestled besides a rustic, wooden crate full of freshly harvested vegetables. |
+
+Table 1. We show examples for the three synthesis templates. Such examples are used as demonstrations for Llama-2 to perform the in-context learning task. We have 176 such examples in total. Most of them are generated by prompting GPT-4 [58], while a handful of others are human generated (in a 10M scale pilot study of synthetic captions, we do not notice significant differences between including or excluding human generated examples.)
+
+
+Figure 3. In-context caption generation using Llama-2 [83]. We randomly sample three in-context examples for each inference run.
+
+of, our objective is to prompt the LLM to create captions such as a cute yellow cat is enjoying the fish in front of the sofa. To add variety, we have a selection of 10 different positional relationship words that we randomly choose from.
+
+For each of the three templates, we have prepared multiple demonstration examples that serve as instructions for the LLM to complete the caption synthesis task. Table 1 shows a couple of examples for each template. In total, we have 106 examples for $c->prompt$ , 50 examples for $c, bg->prompt$ , and 20 examples for $c, rel->prompt$ . Such examples are mostly collected by prompting GPT-4, with a handful from human. In a pilot study, we do not observe difference between including or excluding human generated examples.
+
+In the stage of generating captions in-context, we select a concept and one of the three templates. Next, we randomly pick three examples from the chosen template and frame the caption generation as a text completion task. This process is illustrated in Figure 3.
+
+# 3.2. Synthesizing Images
+
+For each text caption, we generate a variety of images by initiating the reverse diffusion process with different random noise. The Classifier-Free Guidance (CFG) scale is a crucial factor in this process. A higher CFG scale enhances the quality of the samples and the alignment between text and image, whereas a lower scale results in more diverse samples and better adherence to the original conditional distribution of images based on the given text. Following the approach used in StableRep [81], we opt for a lower CFG scale, specifically 2.5, and produce 4 images for each caption. Examples of these images can be seen in Figure 4.
+
+# 3.3. Representation Learning
+
+Our representation learning method is built upon StableRep [81]. The key component of our approach is the multi-positive contrastive learning loss [43] which works by aligning (in the embedding space) images generated from the same caption. We additionally combine multiple techniques from other self-supervised learning methods, including a patch-level masked image modeling objective. We briefly review StableRep and elaborate on the added modules.
+
+StableRep [81] minimizes the cross-entropy loss between a ground-truth assignment distribution and a contrastive assignment distribution. Consider an encoded anchor sample $\pmb{a}$ and a set of encoded candidates $\{\pmb{b}_1, \pmb{b}_2, \dots, \pmb{b}_K\}$ . The contrastive assignment distribution $\mathbf{q}$ describes how likely the model predicts $\pmb{a}$ and each $\pmb{b}$ to be generated from the same caption, and the ground-truth distribution is the actual match
+
+
+A plate of paella, a mixed rice dish with chicken, beans, and seafood
+
+
+An industrial power plant with its smokestacks belching black smoke.
+
+
+A fluffy, black and white junco bird perches on a snow-covered fence, overlooking a dark forest.
+Figure 4. Random examples of synthetic captions and images generated in our SynCLR pipeline. Each caption comes with 4 images.
+
+between $\pmb{a}$ and $\pmb{b}$ ( $\pmb{a}$ is allowed to match multiple $\pmb{b}$ ):
+
+$$
+\mathbf {q} _ {i} = \frac {\exp \left(\boldsymbol {a} \cdot \boldsymbol {b} _ {i} / \tau\right)}{\sum_ {j = 1} ^ {K} \exp \left(\boldsymbol {a} \cdot \boldsymbol {b} _ {j} / \tau\right)} \tag {1}
+$$
+
+$$
+\mathbf {p} _ {i} = \frac {\mathbb {1} _ {\operatorname {m a t c h} \left(\boldsymbol {a} , \boldsymbol {b} _ {i}\right)}}{\sum_ {j = 1} ^ {K} \mathbb {1} _ {\operatorname {m a t c h} \left(\boldsymbol {a} , \boldsymbol {b} _ {j}\right)}} \tag {2}
+$$
+
+where $\tau \in \mathcal{R}_{+}$ is the scalar temperature, $\pmb{a}$ and all $\pmb{b}$ have been $\ell_{2}$ normalized, and the indicator function $\mathbb{1}_{\mathrm{match}(\cdot ,\cdot)}$ indicates whether two samples are from the same caption. The contrastive loss for $\pmb{a}$ is given as
+
+$$
+\mathcal {L} (\boldsymbol {a}) = H (\boldsymbol {p}, \boldsymbol {q}) = - \sum_ {i = 1} ^ {K} \mathbf {p} _ {i} \log \mathbf {q} _ {i} \tag {3}
+$$
+
+iBOT [98] is a masked image modeling objective, wherein a localized patch is masked, and the model is tasked with predicting the tokenized representation of said masked patch. It adapts the DINO [11] objective from the image level into the patch level. We follow [67] to replace the softmax-centering method with the iterative Sinkhorn-Knopp (SK) algorithm [19]. We run SK for 3 iterations to build the prediction target.
+
+Exponential Moving Average (EMA) is firstly introduced into self-supervised learning by MoCo [38]. We use EMA to encode crops as $\mathbf{b}$ and to produce the targets for iBOT loss. We update the EMA model as $\theta_{ema} \gets \lambda \theta_{ema} + (1 - \lambda)\theta$ , following a cosine schedule for $\lambda$ from 0.994 to 1 during training [34, 59]. We find the EMA module not only increases the final performance, but also improves the training stability for long training schedules.
+
+Multi-crop strategy is introduced by [10] as a smart way to improve computation efficiency, and is adopted in this paper.
+
+For these local crops, we only employ the contrastive loss, omitting the iBOT loss. Local crops are encoded only by the student network, and matched to global crops from the same caption encoded by the EMA model. Such reuse of global crops saves computation. For each image $x$ , where we generate a single global crop $x^g$ alongside $n$ local crops $x^l$ , the final loss can be expressed as follows:
+
+$$
+\mathcal {L} \left(x ^ {g}\right) + \frac {1}{n} \sum_ {i = 1} ^ {n} \mathcal {L} \left(x _ {i} ^ {l}\right) + \mathcal {L} ^ {i B O T} \left(x ^ {g}\right) \tag {4}
+$$
+
+# 3.4. Implementation
+
+Concept list. We concatenate class names from various datasets, including IN-1k [21], IN-21k (we keep the most frequent 13k classes), Aircraft [51], Cars [44], DTD [17], Flowers [55], Pets [60], Sun397 [88], Caltech-101 [30], Food101 [7], and Places-365 [96]. If the concept is a place (i.e. SUN397 and Places) or a texture (i.e. DTD), we only apply the $c \rightarrow$ caption template. For fine-grained classes such as pets or flowers, we employ GPT-4 to generate a consolidated list of probable backgrounds, rather than producing distinct lists for each specific class. We favor more frequent sampling from IN-1k, Food101, Cars, Aircraft, and Flowers.
+
+Batches. For each training batch, we sample 2048 captions (except when noted), and use all of the 4 images generated by each caption. We generate 1 global and 4 local crops for each image. As a result, each batch contains 8192 global crops, which is similar with prior work [13, 14, 34, 81].
+
+Masking. For the iBOT loss, we randomly choose $50\%$ images inside a batch to mask, and randomly mask $50\%$ of the tokens in each chosen image. We use 65536 prototypes. While the target from the EMA model is ascertained using the SK algorithm, we apply softmax normalization to the output of the student model.
+
+Projection heads. We follow the design in MoCo v3 [14] and DINO [11] for the contrastive and iBOT loss heads, respectively, ensuring consistency with established methods. Other hyper-parameters. We set the temperature in the contrastive loss to 0.08. For the temperature used in the iBOT loss, we linearly increase it from 0.04 to 0.07 over 4000 iterations, and keep it as 0.07 afterwards, as in DINO [11]. Additionally, the weight decay parameter is incrementally adjusted from 0.04 to 0.2, adhering to a cosine schedule.
+
+# 4. Experiment
+
+We first perform an ablation study to evaluate the efficacy of various designs and modules within our pipeline. Then we proceed to scale up the volume of synthetic data.
+
+# 4.1. Study different components
+
+We analyze each component of SynCLR, and ablate their effectiveness in two measurements: (1) linear probing performance on IN-1k; (2) average accuracy of linear transfer on
+
+| Templates | In context examples |
| c->caption | revolver ->Multiple antique revolvers lie on a wooden table, gleaming under soft, ambient light. closetet ->The compact closet, brimming with clothes and shoes, exudes a feeling of organization. zebra ->A zebra is gallantly trotting across the vast, sunlit plains of the African savannah, creating a captivating black and white spectacle. bus station ->The bustling bus station thrums with restless energy, as travelers navigate through the crowded space, awaiting their journeys amid the echoes of departing buses. |
| c,bg->caption | tiger, forest ->Two tigers are running together in the forest. lighter, motorhome ->In the cozy, cluttered environment of a well-traveled motorhome, a sleek silver lighter holds dominion on the rustic wooden table. sunset, lake ->Golden sunset hues reflect on a calm lake, silhouetting a lone canoeist against a backdrop of fiery clouds. |
| c,rel->caption | kit fox, in front of ->A group of small, fluffy, golden kit foxes is playfully gathered in front of a lush, green, towering forest backdrop. cabbage, besides ->A vibrant image portrays a lush, green cabbage, glistening with dewdrops, nestled besides a rustic, wooden crate full of freshly harvested vegetables. |
",
+ "image_path": "9c059581603d7f922b72bd316a2a38112c5b7468661aa8b6f2c1dd389a6c4564.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 46,
+ 262,
+ 547,
+ 308
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 262,
+ 547,
+ 308
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 262,
+ 547,
+ 308
+ ],
+ "type": "text",
+ "content": "Table 1. We show examples for the three synthesis templates. Such examples are used as demonstrations for Llama-2 to perform the in-context learning task. We have 176 such examples in total. Most of them are generated by prompting GPT-4 [58], while a handful of others are human generated (in a 10M scale pilot study of synthetic captions, we do not notice significant differences between including or excluding human generated examples.)"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "text"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 50,
+ 323,
+ 286,
+ 415
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 50,
+ 323,
+ 286,
+ 415
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 50,
+ 323,
+ 286,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 50,
+ 323,
+ 286,
+ 415
+ ],
+ "type": "image",
+ "image_path": "8a42f1841f22903e4b63eb70e60d711d71f5a05e01cac9bfe87c99dfd12c5267.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 47,
+ 423,
+ 288,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 423,
+ 288,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 423,
+ 288,
+ 447
+ ],
+ "type": "text",
+ "content": "Figure 3. In-context caption generation using Llama-2 [83]. We randomly sample three in-context examples for each inference run."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 55,
+ 471,
+ 287,
+ 531
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 55,
+ 471,
+ 287,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 55,
+ 471,
+ 287,
+ 531
+ ],
+ "type": "text",
+ "content": "of, our objective is to prompt the LLM to create captions such as a cute yellow cat is enjoying the fish in front of the sofa. To add variety, we have a selection of 10 different positional relationship words that we randomly choose from."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "text",
+ "content": "For each of the three templates, we have prepared multiple demonstration examples that serve as instructions for the LLM to complete the caption synthesis task. Table 1 shows a couple of examples for each template. In total, we have 106 examples for "
+ },
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "inline_equation",
+ "content": "c->prompt"
+ },
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "text",
+ "content": ", 50 examples for "
+ },
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "inline_equation",
+ "content": "c, bg->prompt"
+ },
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "text",
+ "content": ", and 20 examples for "
+ },
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "inline_equation",
+ "content": "c, rel->prompt"
+ },
+ {
+ "bbox": [
+ 46,
+ 541,
+ 288,
+ 650
+ ],
+ "type": "text",
+ "content": ". Such examples are mostly collected by prompting GPT-4, with a handful from human. In a pilot study, we do not observe difference between including or excluding human generated examples."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 46,
+ 654,
+ 288,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 654,
+ 288,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 654,
+ 288,
+ 715
+ ],
+ "type": "text",
+ "content": "In the stage of generating captions in-context, we select a concept and one of the three templates. Next, we randomly pick three examples from the chosen template and frame the caption generation as a text completion task. This process is illustrated in Figure 3."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 306,
+ 323,
+ 424,
+ 336
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 323,
+ 424,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 323,
+ 424,
+ 336
+ ],
+ "type": "text",
+ "content": "3.2. Synthesizing Images"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 304,
+ 346,
+ 547,
+ 479
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 346,
+ 547,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 346,
+ 547,
+ 479
+ ],
+ "type": "text",
+ "content": "For each text caption, we generate a variety of images by initiating the reverse diffusion process with different random noise. The Classifier-Free Guidance (CFG) scale is a crucial factor in this process. A higher CFG scale enhances the quality of the samples and the alignment between text and image, whereas a lower scale results in more diverse samples and better adherence to the original conditional distribution of images based on the given text. Following the approach used in StableRep [81], we opt for a lower CFG scale, specifically 2.5, and produce 4 images for each caption. Examples of these images can be seen in Figure 4."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 306,
+ 503,
+ 446,
+ 517
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 503,
+ 446,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 503,
+ 446,
+ 517
+ ],
+ "type": "text",
+ "content": "3.3. Representation Learning"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 304,
+ 528,
+ 547,
+ 624
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 528,
+ 547,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 528,
+ 547,
+ 624
+ ],
+ "type": "text",
+ "content": "Our representation learning method is built upon StableRep [81]. The key component of our approach is the multi-positive contrastive learning loss [43] which works by aligning (in the embedding space) images generated from the same caption. We additionally combine multiple techniques from other self-supervised learning methods, including a patch-level masked image modeling objective. We briefly review StableRep and elaborate on the added modules."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": "StableRep [81] minimizes the cross-entropy loss between a ground-truth assignment distribution and a contrastive assignment distribution. Consider an encoded anchor sample "
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{a}"
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": " and a set of encoded candidates "
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\{\\pmb{b}_1, \\pmb{b}_2, \\dots, \\pmb{b}_K\\}"
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": ". The contrastive assignment distribution "
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{q}"
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": " describes how likely the model predicts "
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{a}"
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": " and each "
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{b}"
+ },
+ {
+ "bbox": [
+ 304,
+ 629,
+ 548,
+ 715
+ ],
+ "type": "text",
+ "content": " to be generated from the same caption, and the ground-truth distribution is the actual match"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "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": "15890"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 61,
+ 70,
+ 274,
+ 125
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 61,
+ 70,
+ 274,
+ 125
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 61,
+ 70,
+ 274,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 61,
+ 70,
+ 274,
+ 125
+ ],
+ "type": "image",
+ "image_path": "dc99af57ce94112007c0043dbe70f56a359cc6068754a28f472dc85053281589.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 0,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 62,
+ 124,
+ 247,
+ 131
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 62,
+ 124,
+ 247,
+ 131
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 62,
+ 124,
+ 247,
+ 131
+ ],
+ "type": "text",
+ "content": "A plate of paella, a mixed rice dish with chicken, beans, and seafood"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 61,
+ 138,
+ 274,
+ 193
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 61,
+ 138,
+ 274,
+ 193
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 61,
+ 138,
+ 274,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 61,
+ 138,
+ 274,
+ 193
+ ],
+ "type": "image",
+ "image_path": "9c549d9d6852c2bed63f35884b2498ecbe23a23b1be85f1137677ab6c2fb8d44.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 62,
+ 193,
+ 249,
+ 199
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 62,
+ 193,
+ 249,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 62,
+ 193,
+ 249,
+ 199
+ ],
+ "type": "text",
+ "content": "An industrial power plant with its smokestacks belching black smoke."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 61,
+ 201,
+ 274,
+ 255
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 61,
+ 201,
+ 274,
+ 255
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 61,
+ 201,
+ 274,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 61,
+ 201,
+ 274,
+ 255
+ ],
+ "type": "image",
+ "image_path": "7183cd8ec8c687dba982a5f8c99be5008b3add07a27b232ead4282a53c4e14f9.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 62,
+ 255,
+ 253,
+ 269
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 62,
+ 255,
+ 253,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 62,
+ 255,
+ 253,
+ 269
+ ],
+ "type": "text",
+ "content": "A fluffy, black and white junco bird perches on a snow-covered fence, overlooking a dark forest."
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 47,
+ 274,
+ 287,
+ 306
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 274,
+ 287,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 274,
+ 287,
+ 306
+ ],
+ "type": "text",
+ "content": "Figure 4. Random examples of synthetic captions and images generated in our SynCLR pipeline. Each caption comes with 4 images."
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "text",
+ "content": "between "
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{a}"
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{b}"
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "text",
+ "content": " ("
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{a}"
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "text",
+ "content": " is allowed to match multiple "
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{b}"
+ },
+ {
+ "bbox": [
+ 47,
+ 320,
+ 258,
+ 333
+ ],
+ "type": "text",
+ "content": "):"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 111,
+ 338,
+ 287,
+ 369
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 338,
+ 287,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 338,
+ 287,
+ 369
+ ],
+ "type": "interline_equation",
+ "content": "\\mathbf {q} _ {i} = \\frac {\\exp \\left(\\boldsymbol {a} \\cdot \\boldsymbol {b} _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\boldsymbol {a} \\cdot \\boldsymbol {b} _ {j} / \\tau\\right)} \\tag {1}",
+ "image_path": "e318576116f5d14fb85e13a74108a882d9d80465b5749327c6e97682058479fb.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 123,
+ 370,
+ 287,
+ 399
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 123,
+ 370,
+ 287,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 123,
+ 370,
+ 287,
+ 399
+ ],
+ "type": "interline_equation",
+ "content": "\\mathbf {p} _ {i} = \\frac {\\mathbb {1} _ {\\operatorname {m a t c h} \\left(\\boldsymbol {a} , \\boldsymbol {b} _ {i}\\right)}}{\\sum_ {j = 1} ^ {K} \\mathbb {1} _ {\\operatorname {m a t c h} \\left(\\boldsymbol {a} , \\boldsymbol {b} _ {j}\\right)}} \\tag {2}",
+ "image_path": "a71eb6addba44bdb80d42d4afe7efc8b5329d5909e0dce5d2d9895c34e1ddd8f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "inline_equation",
+ "content": "\\tau \\in \\mathcal{R}_{+}"
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "content": " is the scalar temperature, "
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{a}"
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "content": " and all "
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{b}"
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "content": " have been "
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "inline_equation",
+ "content": "\\ell_{2}"
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "content": " normalized, and the indicator function "
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbb{1}_{\\mathrm{match}(\\cdot ,\\cdot)}"
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "content": " indicates whether two samples are from the same caption. The contrastive loss for "
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "inline_equation",
+ "content": "\\pmb{a}"
+ },
+ {
+ "bbox": [
+ 47,
+ 405,
+ 287,
+ 453
+ ],
+ "type": "text",
+ "content": " is given as"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 94,
+ 460,
+ 287,
+ 493
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 94,
+ 460,
+ 287,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 94,
+ 460,
+ 287,
+ 493
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} (\\boldsymbol {a}) = H (\\boldsymbol {p}, \\boldsymbol {q}) = - \\sum_ {i = 1} ^ {K} \\mathbf {p} _ {i} \\log \\mathbf {q} _ {i} \\tag {3}",
+ "image_path": "0181ec253f36b888af31a77c3fe1d19cf3f590659149fa22ecf99f7b287be17f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 46,
+ 498,
+ 287,
+ 594
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 498,
+ 287,
+ 594
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 498,
+ 287,
+ 594
+ ],
+ "type": "text",
+ "content": "iBOT [98] is a masked image modeling objective, wherein a localized patch is masked, and the model is tasked with predicting the tokenized representation of said masked patch. It adapts the DINO [11] objective from the image level into the patch level. We follow [67] to replace the softmax-centering method with the iterative Sinkhorn-Knopp (SK) algorithm [19]. We run SK for 3 iterations to build the prediction target."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "text",
+ "content": "Exponential Moving Average (EMA) is firstly introduced into self-supervised learning by MoCo [38]. We use EMA to encode crops as "
+ },
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "inline_equation",
+ "content": "\\mathbf{b}"
+ },
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "text",
+ "content": " and to produce the targets for iBOT loss. We update the EMA model as "
+ },
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "inline_equation",
+ "content": "\\theta_{ema} \\gets \\lambda \\theta_{ema} + (1 - \\lambda)\\theta"
+ },
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "text",
+ "content": ", following a cosine schedule for "
+ },
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "inline_equation",
+ "content": "\\lambda"
+ },
+ {
+ "bbox": [
+ 46,
+ 594,
+ 288,
+ 689
+ ],
+ "type": "text",
+ "content": " from 0.994 to 1 during training [34, 59]. We find the EMA module not only increases the final performance, but also improves the training stability for long training schedules."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 47,
+ 689,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 47,
+ 689,
+ 288,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 47,
+ 689,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": "Multi-crop strategy is introduced by [10] as a smart way to improve computation efficiency, and is adopted in this paper."
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "text",
+ "content": "For these local crops, we only employ the contrastive loss, omitting the iBOT loss. Local crops are encoded only by the student network, and matched to global crops from the same caption encoded by the EMA model. Such reuse of global crops saves computation. For each image "
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "inline_equation",
+ "content": "x"
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "text",
+ "content": ", where we generate a single global crop "
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "inline_equation",
+ "content": "x^g"
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "text",
+ "content": " alongside "
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "inline_equation",
+ "content": "n"
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "text",
+ "content": " local crops "
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "inline_equation",
+ "content": "x^l"
+ },
+ {
+ "bbox": [
+ 304,
+ 72,
+ 547,
+ 156
+ ],
+ "type": "text",
+ "content": ", the final loss can be expressed as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 351,
+ 162,
+ 546,
+ 193
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 351,
+ 162,
+ 546,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 351,
+ 162,
+ 546,
+ 193
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} \\left(x ^ {g}\\right) + \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathcal {L} \\left(x _ {i} ^ {l}\\right) + \\mathcal {L} ^ {i B O T} \\left(x ^ {g}\\right) \\tag {4}",
+ "image_path": "ad1dc0bd08b9b8948bd50fc510b461562426105341e99ef0d14dbfe88eec2ccf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 306,
+ 198,
+ 404,
+ 211
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 198,
+ 404,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 198,
+ 404,
+ 211
+ ],
+ "type": "text",
+ "content": "3.4. Implementation"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 304,
+ 217,
+ 547,
+ 347
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 217,
+ 547,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 217,
+ 547,
+ 347
+ ],
+ "type": "text",
+ "content": "Concept list. We concatenate class names from various datasets, including IN-1k [21], IN-21k (we keep the most frequent 13k classes), Aircraft [51], Cars [44], DTD [17], Flowers [55], Pets [60], Sun397 [88], Caltech-101 [30], Food101 [7], and Places-365 [96]. If the concept is a place (i.e. SUN397 and Places) or a texture (i.e. DTD), we only apply the "
+ },
+ {
+ "bbox": [
+ 304,
+ 217,
+ 547,
+ 347
+ ],
+ "type": "inline_equation",
+ "content": "c \\rightarrow"
+ },
+ {
+ "bbox": [
+ 304,
+ 217,
+ 547,
+ 347
+ ],
+ "type": "text",
+ "content": " caption template. For fine-grained classes such as pets or flowers, we employ GPT-4 to generate a consolidated list of probable backgrounds, rather than producing distinct lists for each specific class. We favor more frequent sampling from IN-1k, Food101, Cars, Aircraft, and Flowers."
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 304,
+ 348,
+ 547,
+ 408
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 348,
+ 547,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 348,
+ 547,
+ 408
+ ],
+ "type": "text",
+ "content": "Batches. For each training batch, we sample 2048 captions (except when noted), and use all of the 4 images generated by each caption. We generate 1 global and 4 local crops for each image. As a result, each batch contains 8192 global crops, which is similar with prior work [13, 14, 34, 81]."
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 304,
+ 407,
+ 547,
+ 479
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 407,
+ 547,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 407,
+ 547,
+ 479
+ ],
+ "type": "text",
+ "content": "Masking. For the iBOT loss, we randomly choose "
+ },
+ {
+ "bbox": [
+ 304,
+ 407,
+ 547,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "50\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 407,
+ 547,
+ 479
+ ],
+ "type": "text",
+ "content": " images inside a batch to mask, and randomly mask "
+ },
+ {
+ "bbox": [
+ 304,
+ 407,
+ 547,
+ 479
+ ],
+ "type": "inline_equation",
+ "content": "50\\%"
+ },
+ {
+ "bbox": [
+ 304,
+ 407,
+ 547,
+ 479
+ ],
+ "type": "text",
+ "content": " of the tokens in each chosen image. We use 65536 prototypes. While the target from the EMA model is ascertained using the SK algorithm, we apply softmax normalization to the output of the student model."
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 304,
+ 479,
+ 547,
+ 587
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 479,
+ 547,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 479,
+ 547,
+ 587
+ ],
+ "type": "text",
+ "content": "Projection heads. We follow the design in MoCo v3 [14] and DINO [11] for the contrastive and iBOT loss heads, respectively, ensuring consistency with established methods. Other hyper-parameters. We set the temperature in the contrastive loss to 0.08. For the temperature used in the iBOT loss, we linearly increase it from 0.04 to 0.07 over 4000 iterations, and keep it as 0.07 afterwards, as in DINO [11]. Additionally, the weight decay parameter is incrementally adjusted from 0.04 to 0.2, adhering to a cosine schedule."
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 306,
+ 597,
+ 382,
+ 610
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 597,
+ 382,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 597,
+ 382,
+ 610
+ ],
+ "type": "text",
+ "content": "4. Experiment"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 304,
+ 616,
+ 546,
+ 653
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 616,
+ 546,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 616,
+ 546,
+ 653
+ ],
+ "type": "text",
+ "content": "We first perform an ablation study to evaluate the efficacy of various designs and modules within our pipeline. Then we proceed to scale up the volume of synthetic data."
+ }
+ ]
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 306,
+ 659,
+ 459,
+ 672
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 306,
+ 659,
+ 459,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 306,
+ 659,
+ 459,
+ 672
+ ],
+ "type": "text",
+ "content": "4.1. Study different components"
+ }
+ ]
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 304,
+ 677,
+ 547,
+ 713
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 304,
+ 677,
+ 547,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 304,
+ 677,
+ 547,
+ 713
+ ],
+ "type": "text",
+ "content": "We analyze each component of SynCLR, and ablate their effectiveness in two measurements: (1) linear probing performance on IN-1k; (2) average accuracy of linear transfer on"
+ }
+ ]
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "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": "15891"
+ }
+ ]
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 73,
+ 60,
+ 261,
+ 155
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 73,
+ 60,
+ 261,
+ 155
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 73,
+ 60,
+ 261,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 73,
+ 60,
+ 261,
+ 155
+ ],
+ "type": "table",
+ "html": "| Layer | Output size |
| conv1 | 3 × 3 × 3, 64, stride(1,1,1) | T × H × W |
| stage1 | [1 × 1 × 1, 64, (0,0,0)FTSCnv, 64, (1,1,1)1 × 1 × 1, 64, (0,0,0)] × 3 | T × H × W |
| stage2 | [1 × 1 × 1, 128, (0,0,0)FTSCnv, 128, (1,2,2)1 × 1 × 1, 128, (0,0,0)] × 4 | T × H/2 × W/2 |
| stage3 | [1 × 1 × 1, 256, (0,0,0)FTSCnv, 256, (1,2,2)1 × 1 × 1, 256, (0,0,0)] × 6 | T × H/4 × W/4 |
| stage4 | [1 × 1 × 1, 512, (0,0,0)FTSCnv, 512, (1,1,1)1 × 1 × 1, 512, (0,0,0)] × 3 | T × H/4 × W/4 |
",
+ "bbox": [
+ 96,
+ 273,
+ 444,
+ 479
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "We evaluate our method on the widely adopted Gait3D [61] and GREW [63] datasets in Tab. 1. To enhance the robustness of our model, we employ three data augmentation techniques: rotation, flipping, and perspective. We conduct training using VPNet-L on both Gait3D and GREW datasets and report rank-1 (\\%) and rank-5 (\\%) accuracy.",
+ "bbox": [
+ 75,
+ 489,
+ 468,
+ 580
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "In-the-lab datasets. In-the-lab datasets encompass various controllable conditions, including camera angles, predetermined walking routes, and dress variations. These datasets play a pivotal role in evaluating the ability of the gait network to handle cross-viewing and cross-dressing scenarios. We select the well-established OUMVLP [44] and CASIA-B [58] datasets to evaluate our approach in Tab. 1. Specifically, the OUMVLP provides an abundance of cross-viewing data, while the CASIA-B offers valuable cross-dressing data. Limited datasets, particularly those obtained from laboratory scenarios, may result in overfitting if the network parameters are excessively large. Therefore, we employ VPNet-M for training on OUMVLP and VPNet-T for training on CASIA-B. We present the rank-1 $(\\%)$ accuracy for all perspectives, excluding the identical-view cases.",
+ "bbox": [
+ 73,
+ 580,
+ 468,
+ 808
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Implementation Details. We establish the network parameters based on the dataset scale and complexity, as outlined in Tab. 1. The batch size is represented as $(S, T)$ , indicating that each mini-batch involves $S$ subjects. For each subject, $T$ sequences are sampled, and within each sequence, a random selection of 20 to 40 frames is chosen following",
+ "bbox": [
+ 75,
+ 810,
+ 470,
+ 901
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "table",
+ "img_path": "images/cccc4d4f625281589c9942ba48b6703f9aa6ef0cc74cde0650c39f33fcecdefa.jpg",
+ "table_caption": [
+ "Table 3. The performance comparisons on Gait3D are reported with rank-1 and rank-5 accuracy (\\%)."
+ ],
+ "table_footnote": [],
+ "table_body": "| Layer | Output size |
| conv1 | 3 × 3 × 3, 64, stride(1,1,1) | T × H × W |
| stage1 | [1 × 1 × 1, 64, (0,0,0)FTSCnv, 64, (1,1,1)1 × 1 × 1, 64, (0,0,0)] × 3 | T × H × W |
| stage2 | [1 × 1 × 1, 128, (0,0,0)FTSCnv, 128, (1,2,2)1 × 1 × 1, 128, (0,0,0)] × 4 | T × H/2 × W/2 |
| stage3 | [1 × 1 × 1, 256, (0,0,0)FTSCnv, 256, (1,2,2)1 × 1 × 1, 256, (0,0,0)] × 6 | T × H/4 × W/4 |
| stage4 | [1 × 1 × 1, 512, (0,0,0)FTSCnv, 512, (1,1,1)1 × 1 × 1, 512, (0,0,0)] × 3 | T × H/4 × W/4 |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.49,
+ 0.47,
+ 0.581
+ ],
+ "angle": 0,
+ "content": "We evaluate our method on the widely adopted Gait3D [61] and GREW [63] datasets in Tab. 1. To enhance the robustness of our model, we employ three data augmentation techniques: rotation, flipping, and perspective. We conduct training using VPNet-L on both Gait3D and GREW datasets and report rank-1 (\\%) and rank-5 (\\%) accuracy."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.075,
+ 0.582,
+ 0.47,
+ 0.809
+ ],
+ "angle": 0,
+ "content": "In-the-lab datasets. In-the-lab datasets encompass various controllable conditions, including camera angles, predetermined walking routes, and dress variations. These datasets play a pivotal role in evaluating the ability of the gait network to handle cross-viewing and cross-dressing scenarios. We select the well-established OUMVLP [44] and CASIA-B [58] datasets to evaluate our approach in Tab. 1. Specifically, the OUMVLP provides an abundance of cross-viewing data, while the CASIA-B offers valuable cross-dressing data. Limited datasets, particularly those obtained from laboratory scenarios, may result in overfitting if the network parameters are excessively large. Therefore, we employ VPNet-M for training on OUMVLP and VPNet-T for training on CASIA-B. We present the rank-1 \\((\\%)\\) accuracy for all perspectives, excluding the identical-view cases."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.076,
+ 0.811,
+ 0.471,
+ 0.902
+ ],
+ "angle": 0,
+ "content": "Implementation Details. We establish the network parameters based on the dataset scale and complexity, as outlined in Tab. 1. The batch size is represented as \\((S, T)\\), indicating that each mini-batch involves \\(S\\) subjects. For each subject, \\(T\\) sequences are sampled, and within each sequence, a random selection of 20 to 40 frames is chosen following"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.499,
+ 0.234,
+ 0.894,
+ 0.264
+ ],
+ "angle": 0,
+ "content": "Table 3. The performance comparisons on Gait3D are reported with rank-1 and rank-5 accuracy (\\%)."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.52,
+ 0.274,
+ 0.871,
+ 0.462
+ ],
+ "angle": 0,
+ "content": "| Layer | Output size |
| conv1 | 3 × 3 × 3, 64, stride(1,1,1) | T × H × W |
| stage1 | [1 × 1 × 1, 64, (0,0,0)FTSCnv, 64, (1,1,1)1 × 1 × 1, 64, (0,0,0)] × 3 | T × H × W |
| stage2 | [1 × 1 × 1, 128, (0,0,0)FTSCnv, 128, (1,2,2)1 × 1 × 1, 128, (0,0,0)] × 4 | T × H/2 × W/2 |
| stage3 | [1 × 1 × 1, 256, (0,0,0)FTSCnv, 256, (1,2,2)1 × 1 × 1, 256, (0,0,0)] × 6 | T × H/4 × W/4 |
| stage4 | [1 × 1 × 1, 512, (0,0,0)FTSCnv, 512, (1,1,1)1 × 1 × 1, 512, (0,0,0)] × 3 | T × H/4 × W/4 |
+
+We evaluate our method on the widely adopted Gait3D [61] and GREW [63] datasets in Tab. 1. To enhance the robustness of our model, we employ three data augmentation techniques: rotation, flipping, and perspective. We conduct training using VPNet-L on both Gait3D and GREW datasets and report rank-1 (\%) and rank-5 (\%) accuracy.
+
+In-the-lab datasets. In-the-lab datasets encompass various controllable conditions, including camera angles, predetermined walking routes, and dress variations. These datasets play a pivotal role in evaluating the ability of the gait network to handle cross-viewing and cross-dressing scenarios. We select the well-established OUMVLP [44] and CASIA-B [58] datasets to evaluate our approach in Tab. 1. Specifically, the OUMVLP provides an abundance of cross-viewing data, while the CASIA-B offers valuable cross-dressing data. Limited datasets, particularly those obtained from laboratory scenarios, may result in overfitting if the network parameters are excessively large. Therefore, we employ VPNet-M for training on OUMVLP and VPNet-T for training on CASIA-B. We present the rank-1 $(\%)$ accuracy for all perspectives, excluding the identical-view cases.
+
+Implementation Details. We establish the network parameters based on the dataset scale and complexity, as outlined in Tab. 1. The batch size is represented as $(S, T)$ , indicating that each mini-batch involves $S$ subjects. For each subject, $T$ sequences are sampled, and within each sequence, a random selection of 20 to 40 frames is chosen following
+
+Table 3. The performance comparisons on Gait3D are reported with rank-1 and rank-5 accuracy (\%).
+
+| Layer | Output size |
| conv1 | 3 × 3 × 3, 64, stride(1,1,1) | T × H × W |
| stage1 | [1 × 1 × 1, 64, (0,0,0)FTSCnv, 64, (1,1,1)1 × 1 × 1, 64, (0,0,0)] × 3 | T × H × W |
| stage2 | [1 × 1 × 1, 128, (0,0,0)FTSCnv, 128, (1,2,2)1 × 1 × 1, 128, (0,0,0)] × 4 | T × H/2 × W/2 |
| stage3 | [1 × 1 × 1, 256, (0,0,0)FTSCnv, 256, (1,2,2)1 × 1 × 1, 256, (0,0,0)] × 6 | T × H/4 × W/4 |
| stage4 | [1 × 1 × 1, 512, (0,0,0)FTSCnv, 512, (1,1,1)1 × 1 × 1, 512, (0,0,0)] × 3 | T × H/4 × W/4 |
",
+ "image_path": "b6fb057be70ca0228df6bc2110d5607822e771086d52487d06029d7cc4d206d3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 46,
+ 388,
+ 287,
+ 460
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 388,
+ 287,
+ 460
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 388,
+ 287,
+ 460
+ ],
+ "type": "text",
+ "content": "We evaluate our method on the widely adopted Gait3D [61] and GREW [63] datasets in Tab. 1. To enhance the robustness of our model, we employ three data augmentation techniques: rotation, flipping, and perspective. We conduct training using VPNet-L on both Gait3D and GREW datasets and report rank-1 (\\%) and rank-5 (\\%) accuracy."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 45,
+ 460,
+ 287,
+ 640
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 45,
+ 460,
+ 287,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 45,
+ 460,
+ 287,
+ 640
+ ],
+ "type": "text",
+ "content": "In-the-lab datasets. In-the-lab datasets encompass various controllable conditions, including camera angles, predetermined walking routes, and dress variations. These datasets play a pivotal role in evaluating the ability of the gait network to handle cross-viewing and cross-dressing scenarios. We select the well-established OUMVLP [44] and CASIA-B [58] datasets to evaluate our approach in Tab. 1. Specifically, the OUMVLP provides an abundance of cross-viewing data, while the CASIA-B offers valuable cross-dressing data. Limited datasets, particularly those obtained from laboratory scenarios, may result in overfitting if the network parameters are excessively large. Therefore, we employ VPNet-M for training on OUMVLP and VPNet-T for training on CASIA-B. We present the rank-1 "
+ },
+ {
+ "bbox": [
+ 45,
+ 460,
+ 287,
+ 640
+ ],
+ "type": "inline_equation",
+ "content": "(\\%)"
+ },
+ {
+ "bbox": [
+ 45,
+ 460,
+ 287,
+ 640
+ ],
+ "type": "text",
+ "content": " accuracy for all perspectives, excluding the identical-view cases."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": "Implementation Details. We establish the network parameters based on the dataset scale and complexity, as outlined in Tab. 1. The batch size is represented as "
+ },
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "(S, T)"
+ },
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": ", indicating that each mini-batch involves "
+ },
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "S"
+ },
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": " subjects. For each subject, "
+ },
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "inline_equation",
+ "content": "T"
+ },
+ {
+ "bbox": [
+ 46,
+ 642,
+ 288,
+ 714
+ ],
+ "type": "text",
+ "content": " sequences are sampled, and within each sequence, a random selection of 20 to 40 frames is chosen following"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 318,
+ 217,
+ 533,
+ 365
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 305,
+ 185,
+ 547,
+ 209
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 305,
+ 185,
+ 547,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 305,
+ 185,
+ 547,
+ 209
+ ],
+ "type": "text",
+ "content": "Table 3. The performance comparisons on Gait3D are reported with rank-1 and rank-5 accuracy (\\%)."
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 318,
+ 217,
+ 533,
+ 365
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 318,
+ 217,
+ 533,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 318,
+ 217,
+ 533,
+ 365
+ ],
+ "type": "table",
+ "html": "