| Method | Loss function | Regularization |
| SimCLR, MoCo | ℓ†NT-Xent = ∑Cc=1C(-1/τic+ sim(zic, zi+c)+ 1/τic- max n=1,...,N sim(zi, zi-c)+ βΩ(τic+)-βΩ(τic-)) | (2.1) |
| SimSiam | ℓ†SymNegCos = ∑Cc=1C(-1/2τic+ sim(zi, i+|sg) - 1/2τic+ sim(zi, i+|h, i|sg) + βΩ(τic+)+ βΩ(τic+)) | (2.2) |
| Barlow Twins | ℓ†Cross-Corr = ∑Cc=1C(d' (1-1/τic+ Cll)2+ λ∑l=1d'∑m≠l d' 1/τic- C2lm + β∑l=1d' 1/τic- d' Ω(τic-)-β∑l=1d' 1/τic- d' Ω(τic-)) | (2.3) |
| LGP, CAN | ℓ†InfoNCE = ∑Cc=1C(-1/τic+ sim(zi, ic+)+ 1/τic- max n=1,...,N+1 sim(zi, ic+±) + βΩ(τic+)-βΩ(τic+)) | (2.4) |
",
+ "bbox": [
+ 215,
+ 178,
+ 787,
+ 296
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "of our AMCL loss for various contrastive learning frameworks. The $\\ddagger$ in Eq. (2.1)-(2.4) indicates that these losses represent our multi-head loss versions.",
+ "bbox": [
+ 212,
+ 308,
+ 782,
+ 338
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "In Eq. (2.1), the asterisk $*$ in $\\tau_{in^*}^{c-}$ indicates the index $n^* = \\arg \\max_{n=1,\\dots,N} \\sin(\\boldsymbol{z}_i, \\boldsymbol{z}_{in}^-)$ .",
+ "bbox": [
+ 241,
+ 339,
+ 800,
+ 364
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "In Eq. (2.2), $\\tau_{i}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{i}^{c}), \\phi([\\boldsymbol{h}_{i}^{+}]_{\\mathrm{sg}}) \\rangle\\big)$ and $\\tau_{i}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{i}^{c+}), \\phi([\\boldsymbol{h}_{i}]_{\\mathrm{sg}}) \\rangle\\big)$ . In Eq. (2.3), temperatures are formed as $\\tau_{l}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{l:}^{c}), \\phi(\\boldsymbol{z}_{l:}^{c+}) \\rangle\\big)$ and $\\tau_{lm}^{c-} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{l:}^{c}), \\phi(\\boldsymbol{z}_{m:}^{c+}) \\rangle\\big)$ where $l \\neq m$ and operator ':' simply indexes and concatenates variable as $\\boldsymbol{z}_{l} = [z_{l1}, \\dots, z_{lN}]^{T}$ . In Eq. (2.4), $\\boldsymbol{z}_{in}^{c\\pm} = \\boldsymbol{z}_{in}^{c-}$ and $\\tau_{in}^{c\\pm} = \\tau_{in}^{c-}$ for $n = 1, \\dots, N$ , and $\\boldsymbol{z}_{in}^{c\\pm} = \\boldsymbol{z}_{in}^{c+}$ and $\\tau_{in}^{c\\pm} = \\tau_{i}^{c+}$ if $n = N + 1$ . Finally, notice that for NT-Xent in Eq. (2.1), we have:",
+ "bbox": [
+ 212,
+ 364,
+ 785,
+ 460
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\begin{array}{l} \\frac {1}{\\tau_ {i n ^ {*}} ^ {c -}} \\max _ {n = 1, \\dots , N} \\sin (\\pmb {z} _ {i} ^ {c}, \\pmb {z} _ {i n} ^ {c -}) - \\varOmega (\\tau_ {i n ^ {*}} ^ {c -}) - (2 \\pi) ^ {d ^ {\\prime} / 2} \\\\ \\approx \\log \\sum_ {n = 1} ^ {N} \\frac {1}{(2 \\pi) ^ {d ^ {\\prime} / 2} \\left(\\tau_ {i n} ^ {c -}\\right) ^ {d ^ {\\prime} / 2}} \\exp \\left(\\frac {1}{\\tau_ {i n} ^ {c -}} \\left(\\operatorname {s i m} \\left(z _ {i} ^ {c}, z _ {i n} ^ {c -}\\right) - 1\\right)\\right). \\tag {5} \\\\ \\end{array}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 299,
+ 473,
+ 785,
+ 550
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "The same approximation (with $z_{in}^{c\\pm}$ in place of $z_{in}^{c-}$ ) holds for InfoNCE in Eq. (2.4). Using maximum in Eq. (5), Eq. (2.1) and Eq. (2.4) are somewhat restrictive as the soft-maximum, depending on the temperature, will return the maximum similarity or interpolation over top few similarities close to maximum. Thus, softmax will tackle a group of negative samples closest to the anchor. Indeed, we could use the above soft-maximum in place of maximum but then we have no easy way of recovering the temperature $\\tau_{in*}^{c-}$ . Thus, in our experiments we observed that the best choice is to apply $\\sum_{k=1}^{\\kappa} \\frac{1}{\\kappa \\tau_{in_k}^{c-}} \\left[ \\operatorname{TopMax}_\\kappa \\sin(\\pmb{z}_i^c, \\pmb{z}_{in}^{c-}) \\right]_k$ which is the average over top- $\\kappa$ largest similarities. As the top- $\\kappa$ maximum operation returns also corresponding indexes $n_1^*, \\dots, n_\\kappa^*$ , the temperature regularization can be easily computed as $\\Omega(\\{n_k^*\\}_{k=1}^\\kappa) = \\log(\\tau_{in_1^*}^{c-} \\dots \\tau_{in_\\kappa^*}^{c-}) + \\sum_{k=1}^{\\kappa} \\frac{1}{\\tau_{in_k}^{c-}}$ . The derivation of our loss function is presented in Appendix A.",
+ "bbox": [
+ 212,
+ 561,
+ 787,
+ 762
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "4.4 Discussion",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 782,
+ 349,
+ 797
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Multiple projection heads vs. a wider MLP projection head. Multiple projection heads offer a significant advantage by creating different sets of",
+ "bbox": [
+ 212,
+ 809,
+ 787,
+ 840
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "header",
+ "text": "Adaptive Multi-head Contrastive Learning",
+ "bbox": [
+ 446,
+ 114,
+ 732,
+ 128
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "7",
+ "bbox": [
+ 774,
+ 116,
+ 784,
+ 126
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/6079c683a06bc7c366d37dda567ba0a16a2d58164fb00e7442cba050bbdcad93.jpg",
+ "image_caption": [
+ "(a) head sim. (baseline)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 217,
+ 143,
+ 359,
+ 220
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/3b8f6cdf33cb7fe9f4e8c9895538d18a154551376b6608265dbd13c474460bb0.jpg",
+ "image_caption": [
+ "(b) head sim. (ours)",
+ "Fig. 3: Distribution of similarity scores for positive and negative pairs. The baseline uses one projection head and constant temperature, while our method has multiple projection heads and adaptive temperature. We use SimCLR for pre-training with ResNet-18 on STL-10. After pre-training, we choose 500 positive pairs and 500 negative pairs from the validation to compute the cosine similarity. In (a) and (b), similarity score (temperature scaled) is computed between the 128-dim features extracted from the projection head(s). In (c) and (d), cosine similarity score is computed between the 512-dim features extracted from the backbone after removing the project heads."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 364,
+ 143,
+ 500,
+ 220
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/935b4b8f60ce8c4e01b1e2d5cebb32fbbeb7af7177988b0b60e8b8c98dc30fb3.jpg",
+ "image_caption": [
+ "(c) feat. sim. (baseline)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 501,
+ 143,
+ 638,
+ 220
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "image",
+ "img_path": "images/26e08683f10278536ebf85eaca98a8368a749a9babfc9e34f4d32a3c0a53a66c.jpg",
+ "image_caption": [
+ "(d) feat. sim. (ours)"
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 638,
+ 143,
+ 779,
+ 220
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "features for each pair of images [43]. This helps build a varied representation that captures various aspects of the input, making the model more resilient and adaptable to different transformations. Furthermore, each projection head's ability to specialize in learning specific patterns or features becomes especially useful when dealing with diverse augmentations that require focused attention on distinct content within the images. In the context of contrastive learning, where diverse augmentation strategies and varying intra-sample similarity introduce more variability in positive pairs, using multiple projection heads excels. This approach effectively manages the resulting variability by enabling the learning of diverse representations. As a result, the model becomes skilled at navigating the intricacies of the data, demonstrating its adaptability to the complex nature of the input. On the other hand, the wider MLP projection head takes a distinct approach by consolidating information from various aspects of the input into a single representation. This proves advantageous when a thorough understanding of the data is crucial, and the relationships between features play a vital role [9].",
+ "bbox": [
+ 212,
+ 366,
+ 787,
+ 594
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Visualization of pair similarity distributions. In Fig. 3 we draw the similarity distributions of negative pairs and positive pairs, under the baseline (1 head + constant temperature) and our method (multiple heads + adaptive temperature). When we use the average similarity across the output from the multiple heads, shown in Fig. 3(a) and (b), we can clearly observe better separability brought by our method. It indicates that our method allows for more effective similarity learning of the positive and negative pairs. On the other hand, if we compute the cosine similarity between features extracted right after the backbone, shown in Fig. 3(c) and (d), better separability can again be observed. It illustrates that better similarity learning further benefits representation learning, finally leading to improved linear probing performance.",
+ "bbox": [
+ 212,
+ 595,
+ 787,
+ 762
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Connecting our loss function to existing contrastive learning methods. Our loss function consists of three terms: (i) positive temperature-weighted similarities for positive pairs (ii) negative temperature-weighted similarities for negative pairs, and (iii) a regularization term for positive and negative temperatures. As identified in previous works [44, 47], the alignment (closeness) of features",
+ "bbox": [
+ 212,
+ 763,
+ 787,
+ 842
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "page_number",
+ "text": "8",
+ "bbox": [
+ 217,
+ 114,
+ 228,
+ 126
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "header",
+ "text": "Lei Wang, Piotr Koniusz, Tom Gedeon, and Liang Zheng",
+ "bbox": [
+ 271,
+ 113,
+ 694,
+ 130
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "from positive pairs and the uniformity of the induced distribution of the (normalized) features on the hypersphere are the two key properties in contrastive loss. Our loss function also optimizes these properties and further improves the contrastive learning performance through parameterized pair-wise temperature via re-weighting the positive and negative similarities. Our loss function is a more general form, and when we set the temperature to be a global constant, the constant regularization term no longer affects optimization, and thus the loss function reduces to the traditional contrastive loss.",
+ "bbox": [
+ 212,
+ 146,
+ 787,
+ 263
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "Physical meaning of the regularization term $\\Omega(\\cdot)$ in Eq. (3). The regularization term consists of regularizing both the positive and negative temperatures. During optimisation, the log $\\tau$ term encourages lower positive temperatures and higher negative temperatures, whereas the reverse function term $\\frac{1}{\\tau}$ is in favour of higher positive temperatures and lower negative temperatures. Hence this regularization term balances the learning of positive and negative temperatures. As we jointly optimize network parameters and the temperature via MLE, this problem naturally becomes decomposed in the maximization of similarities for positive pairs (minimization for negative pairs) weighted by the temperature. However, if $\\tau$ was to reach 0 for positive pairs, one would attain a trivial solution. $\\Omega(\\cdot)$ prevents that trivial solution. Intuitively, one can be very certain in similarity of sample pair but there is a price to pay for that certainty, imposed by $\\Omega(\\cdot)$ resulting from the Welsch function. This $\\Omega(\\cdot)$ expresses the prior belief or preference on temperatures. We provide its mathematical insights including connecting temperature to uncertainty in Appendix A. More discussions can be found in Appendix B",
+ "bbox": [
+ 212,
+ 266,
+ 789,
+ 510
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "5 Experiments",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 532,
+ 375,
+ 550
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "We choose popular datasets that are widely used in evaluating the SSL models, including CIFAR-10, CIFAR-100, STL-10, Tiny-[ImageNet and ImageNet. The dataset details are provided in Appendix Below, we describe the experimental setup and evaluations.",
+ "bbox": [
+ 212,
+ 564,
+ 787,
+ 625
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "5.1 Setup",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 648,
+ 310,
+ 664
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "We conduct experiments on the aforementioned datasets following practices outlined by [25]. We consider five different types of transformations for data augmentations: random cropping, random Gaussian blur, color dropping (i.e., randomly converting images to grayscale), color distortion, and random horizontal flipping. For pre-training, we employ Resnet-18 (R18) [22] for SimCLR [9], MoCo [21], SimSiam [10], and Barlow Twins [51] on CIFAR-10, CIFAR-100 and STL-10. For Tiny-ImageNet and ImageNet, we use Resnet-50 (R50) variants, ViT-B and ViT-L. Other settings, such as the architecture of the projection head, remain consistent with the original algorithm configurations. We select $C$ projection heads in range from 2 to 6. We set the temperature bounds ( $\\eta$ and $\\iota$ of sigmoid) in range $[1e - 5, 2]$ on smaller datasets and $[1e - 5, 5]$ for ImageNet and",
+ "bbox": [
+ 212,
+ 672,
+ 787,
+ 843
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "header",
+ "text": "Adaptive Multi-head Contrastive Learning",
+ "bbox": [
+ 446,
+ 114,
+ 732,
+ 128
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "page_number",
+ "text": "9",
+ "bbox": [
+ 774,
+ 116,
+ 785,
+ 126
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "table",
+ "img_path": "images/d794754ae92c94ccbaf3f7cb1646f7f148f3e77e74963546f7849e5a4a4bfd87.jpg",
+ "table_caption": [
+ "Table 3: Impact of AMCL when applied to popular and state-of-the-art SSL methods. On CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet, models are first pretrained for 1,000, 1,000, 800, and 100 epochs, respectively, and then evaluated using linear probing. Backbones are highlighted. All reported results are averages obtained from 5 runs on each dataset for each method (including baselines)."
+ ],
+ "table_footnote": [],
+ "table_body": "| Method | Loss function | Regularization |
| SimCLR, MoCo | ℓ†NT-Xent = ∑Cc=1C(-1/τic+ sim(zic, zi+c)+ 1/τic- max n=1,...,N sim(zi, zi-c)+ βΩ(τic+)-βΩ(τic-)) | (2.1) |
| SimSiam | ℓ†SymNegCos = ∑Cc=1C(-1/2τic+ sim(zi, i+|sg) - 1/2τic+ sim(zi, i+|h, i|sg) + βΩ(τic+)+ βΩ(τic+)) | (2.2) |
| Barlow Twins | ℓ†Cross-Corr = ∑Cc=1C(d' (1-1/τic+ Cll)2+ λ∑l=1d'∑m≠l d' 1/τic- C2lm + β∑l=1d' 1/τic- d' Ω(τic-)-β∑l=1d' 1/τic- d' Ω(τic-)) | (2.3) |
| LGP, CAN | ℓ†InfoNCE = ∑Cc=1C(-1/τic+ sim(zi, ic+)+ 1/τic- max n=1,...,N+1 sim(zi, ic+±) + βΩ(τic+)-βΩ(τic+)) | (2.4) |
"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.309,
+ 0.784,
+ 0.339
+ ],
+ "angle": 0,
+ "content": "of our AMCL loss for various contrastive learning frameworks. The \\(\\ddagger\\) in Eq. (2.1)-(2.4) indicates that these losses represent our multi-head loss versions."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.242,
+ 0.34,
+ 0.802,
+ 0.365
+ ],
+ "angle": 0,
+ "content": "In Eq. (2.1), the asterisk \\(*\\) in \\(\\tau_{in^*}^{c-}\\) indicates the index \\(n^* = \\arg \\max_{n=1,\\dots,N} \\sin(\\boldsymbol{z}_i, \\boldsymbol{z}_{in}^-)\\)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.365,
+ 0.787,
+ 0.462
+ ],
+ "angle": 0,
+ "content": "In Eq. (2.2), \\(\\tau_{i}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{i}^{c}), \\phi([\\boldsymbol{h}_{i}^{+}]_{\\mathrm{sg}}) \\rangle\\big)\\) and \\(\\tau_{i}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{i}^{c+}), \\phi([\\boldsymbol{h}_{i}]_{\\mathrm{sg}}) \\rangle\\big)\\). In Eq. (2.3), temperatures are formed as \\(\\tau_{l}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{l:}^{c}), \\phi(\\boldsymbol{z}_{l:}^{c+}) \\rangle\\big)\\) and \\(\\tau_{lm}^{c-} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{l:}^{c}), \\phi(\\boldsymbol{z}_{m:}^{c+}) \\rangle\\big)\\) where \\(l \\neq m\\) and operator ':' simply indexes and concatenates variable as \\(\\boldsymbol{z}_{l} = [z_{l1}, \\dots, z_{lN}]^{T}\\). In Eq. (2.4), \\(\\boldsymbol{z}_{in}^{c\\pm} = \\boldsymbol{z}_{in}^{c-}\\) and \\(\\tau_{in}^{c\\pm} = \\tau_{in}^{c-}\\) for \\(n = 1, \\dots, N\\), and \\(\\boldsymbol{z}_{in}^{c\\pm} = \\boldsymbol{z}_{in}^{c+}\\) and \\(\\tau_{in}^{c\\pm} = \\tau_{i}^{c+}\\) if \\(n = N + 1\\). Finally, notice that for NT-Xent in Eq. (2.1), we have:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.3,
+ 0.474,
+ 0.786,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "\\[\n\\begin{array}{l} \\frac {1}{\\tau_ {i n ^ {*}} ^ {c -}} \\max _ {n = 1, \\dots , N} \\sin (\\pmb {z} _ {i} ^ {c}, \\pmb {z} _ {i n} ^ {c -}) - \\varOmega (\\tau_ {i n ^ {*}} ^ {c -}) - (2 \\pi) ^ {d ^ {\\prime} / 2} \\\\ \\approx \\log \\sum_ {n = 1} ^ {N} \\frac {1}{(2 \\pi) ^ {d ^ {\\prime} / 2} \\left(\\tau_ {i n} ^ {c -}\\right) ^ {d ^ {\\prime} / 2}} \\exp \\left(\\frac {1}{\\tau_ {i n} ^ {c -}} \\left(\\operatorname {s i m} \\left(z _ {i} ^ {c}, z _ {i n} ^ {c -}\\right) - 1\\right)\\right). \\tag {5} \\\\ \\end{array}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.562,
+ 0.788,
+ 0.763
+ ],
+ "angle": 0,
+ "content": "The same approximation (with \\( z_{in}^{c\\pm} \\) in place of \\( z_{in}^{c-} \\)) holds for InfoNCE in Eq. (2.4). Using maximum in Eq. (5), Eq. (2.1) and Eq. (2.4) are somewhat restrictive as the soft-maximum, depending on the temperature, will return the maximum similarity or interpolation over top few similarities close to maximum. Thus, softmax will tackle a group of negative samples closest to the anchor. Indeed, we could use the above soft-maximum in place of maximum but then we have no easy way of recovering the temperature \\( \\tau_{in*}^{c-} \\). Thus, in our experiments we observed that the best choice is to apply \\( \\sum_{k=1}^{\\kappa} \\frac{1}{\\kappa \\tau_{in_k}^{c-}} \\left[ \\operatorname{TopMax}_\\kappa \\sin(\\pmb{z}_i^c, \\pmb{z}_{in}^{c-}) \\right]_k \\) which is the average over top-\\( \\kappa \\) largest similarities. As the top-\\( \\kappa \\) maximum operation returns also corresponding indexes \\( n_1^*, \\dots, n_\\kappa^* \\), the temperature regularization can be easily computed as \\( \\Omega(\\{n_k^*\\}_{k=1}^\\kappa) = \\log(\\tau_{in_1^*}^{c-} \\dots \\tau_{in_\\kappa^*}^{c-}) + \\sum_{k=1}^{\\kappa} \\frac{1}{\\tau_{in_k}^{c-}} \\). The derivation of our loss function is presented in Appendix A."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.784,
+ 0.35,
+ 0.798
+ ],
+ "angle": 0,
+ "content": "4.4 Discussion"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.81,
+ 0.789,
+ 0.842
+ ],
+ "angle": 0,
+ "content": "Multiple projection heads vs. a wider MLP projection head. Multiple projection heads offer a significant advantage by creating different sets of"
+ }
+ ],
+ [
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.218,
+ 0.116,
+ 0.23,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "8"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.272,
+ 0.114,
+ 0.695,
+ 0.131
+ ],
+ "angle": 0,
+ "content": "Lei Wang, Piotr Koniusz, Tom Gedeon, and Liang Zheng"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.218,
+ 0.144,
+ 0.36,
+ 0.221
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.22,
+ 0.223,
+ 0.359,
+ 0.234
+ ],
+ "angle": 0,
+ "content": "(a) head sim. (baseline)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.365,
+ 0.144,
+ 0.501,
+ 0.221
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.371,
+ 0.223,
+ 0.489,
+ 0.235
+ ],
+ "angle": 0,
+ "content": "(b) head sim. (ours)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.503,
+ 0.145,
+ 0.639,
+ 0.221
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.502,
+ 0.223,
+ 0.638,
+ 0.235
+ ],
+ "angle": 0,
+ "content": "(c) feat. sim. (baseline)"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.64,
+ 0.145,
+ 0.78,
+ 0.221
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.651,
+ 0.223,
+ 0.767,
+ 0.235
+ ],
+ "angle": 0,
+ "content": "(d) feat. sim. (ours)"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.214,
+ 0.239,
+ 0.788,
+ 0.351
+ ],
+ "angle": 0,
+ "content": "Fig. 3: Distribution of similarity scores for positive and negative pairs. The baseline uses one projection head and constant temperature, while our method has multiple projection heads and adaptive temperature. We use SimCLR for pre-training with ResNet-18 on STL-10. After pre-training, we choose 500 positive pairs and 500 negative pairs from the validation to compute the cosine similarity. In (a) and (b), similarity score (temperature scaled) is computed between the 128-dim features extracted from the projection head(s). In (c) and (d), cosine similarity score is computed between the 512-dim features extracted from the backbone after removing the project heads."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.367,
+ 0.789,
+ 0.595
+ ],
+ "angle": 0,
+ "content": "features for each pair of images [43]. This helps build a varied representation that captures various aspects of the input, making the model more resilient and adaptable to different transformations. Furthermore, each projection head's ability to specialize in learning specific patterns or features becomes especially useful when dealing with diverse augmentations that require focused attention on distinct content within the images. In the context of contrastive learning, where diverse augmentation strategies and varying intra-sample similarity introduce more variability in positive pairs, using multiple projection heads excels. This approach effectively manages the resulting variability by enabling the learning of diverse representations. As a result, the model becomes skilled at navigating the intricacies of the data, demonstrating its adaptability to the complex nature of the input. On the other hand, the wider MLP projection head takes a distinct approach by consolidating information from various aspects of the input into a single representation. This proves advantageous when a thorough understanding of the data is crucial, and the relationships between features play a vital role [9]."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.596,
+ 0.789,
+ 0.763
+ ],
+ "angle": 0,
+ "content": "Visualization of pair similarity distributions. In Fig. 3 we draw the similarity distributions of negative pairs and positive pairs, under the baseline (1 head + constant temperature) and our method (multiple heads + adaptive temperature). When we use the average similarity across the output from the multiple heads, shown in Fig. 3(a) and (b), we can clearly observe better separability brought by our method. It indicates that our method allows for more effective similarity learning of the positive and negative pairs. On the other hand, if we compute the cosine similarity between features extracted right after the backbone, shown in Fig. 3(c) and (d), better separability can again be observed. It illustrates that better similarity learning further benefits representation learning, finally leading to improved linear probing performance."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.765,
+ 0.789,
+ 0.843
+ ],
+ "angle": 0,
+ "content": "Connecting our loss function to existing contrastive learning methods. Our loss function consists of three terms: (i) positive temperature-weighted similarities for positive pairs (ii) negative temperature-weighted similarities for negative pairs, and (iii) a regularization term for positive and negative temperatures. As identified in previous works [44, 47], the alignment (closeness) of features"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.447,
+ 0.115,
+ 0.733,
+ 0.13
+ ],
+ "angle": 0,
+ "content": "Adaptive Multi-head Contrastive Learning"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.775,
+ 0.117,
+ 0.787,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "9"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.147,
+ 0.788,
+ 0.265
+ ],
+ "angle": 0,
+ "content": "from positive pairs and the uniformity of the induced distribution of the (normalized) features on the hypersphere are the two key properties in contrastive loss. Our loss function also optimizes these properties and further improves the contrastive learning performance through parameterized pair-wise temperature via re-weighting the positive and negative similarities. Our loss function is a more general form, and when we set the temperature to be a global constant, the constant regularization term no longer affects optimization, and thus the loss function reduces to the traditional contrastive loss."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.267,
+ 0.79,
+ 0.511
+ ],
+ "angle": 0,
+ "content": "Physical meaning of the regularization term \\(\\Omega(\\cdot)\\) in Eq. (3). The regularization term consists of regularizing both the positive and negative temperatures. During optimisation, the log \\(\\tau\\) term encourages lower positive temperatures and higher negative temperatures, whereas the reverse function term \\(\\frac{1}{\\tau}\\) is in favour of higher positive temperatures and lower negative temperatures. Hence this regularization term balances the learning of positive and negative temperatures. As we jointly optimize network parameters and the temperature via MLE, this problem naturally becomes decomposed in the maximization of similarities for positive pairs (minimization for negative pairs) weighted by the temperature. However, if \\(\\tau\\) was to reach 0 for positive pairs, one would attain a trivial solution. \\(\\Omega(\\cdot)\\) prevents that trivial solution. Intuitively, one can be very certain in similarity of sample pair but there is a price to pay for that certainty, imposed by \\(\\Omega(\\cdot)\\) resulting from the Welsch function. This \\(\\Omega(\\cdot)\\) expresses the prior belief or preference on temperatures. We provide its mathematical insights including connecting temperature to uncertainty in Appendix A. More discussions can be found in Appendix B"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.534,
+ 0.377,
+ 0.551
+ ],
+ "angle": 0,
+ "content": "5 Experiments"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.565,
+ 0.788,
+ 0.626
+ ],
+ "angle": 0,
+ "content": "We choose popular datasets that are widely used in evaluating the SSL models, including CIFAR-10, CIFAR-100, STL-10, Tiny-[ImageNet and ImageNet. The dataset details are provided in Appendix Below, we describe the experimental setup and evaluations."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.649,
+ 0.312,
+ 0.665
+ ],
+ "angle": 0,
+ "content": "5.1 Setup"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.674,
+ 0.789,
+ 0.844
+ ],
+ "angle": 0,
+ "content": "We conduct experiments on the aforementioned datasets following practices outlined by [25]. We consider five different types of transformations for data augmentations: random cropping, random Gaussian blur, color dropping (i.e., randomly converting images to grayscale), color distortion, and random horizontal flipping. For pre-training, we employ Resnet-18 (R18) [22] for SimCLR [9], MoCo [21], SimSiam [10], and Barlow Twins [51] on CIFAR-10, CIFAR-100 and STL-10. For Tiny-ImageNet and ImageNet, we use Resnet-50 (R50) variants, ViT-B and ViT-L. Other settings, such as the architecture of the projection head, remain consistent with the original algorithm configurations. We select \\(C\\) projection heads in range from 2 to 6. We set the temperature bounds (\\(\\eta\\) and \\(\\iota\\) of sigmoid) in range \\([1e - 5, 2]\\) on smaller datasets and \\([1e - 5, 5]\\) for ImageNet and"
+ }
+ ],
+ [
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.218,
+ 0.116,
+ 0.236,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "10"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.271,
+ 0.114,
+ 0.695,
+ 0.131
+ ],
+ "angle": 0,
+ "content": "Lei Wang, Piotr Koniusz, Tom Gedeon, and Liang Zheng"
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.214,
+ 0.145,
+ 0.788,
+ 0.215
+ ],
+ "angle": 0,
+ "content": "Table 3: Impact of AMCL when applied to popular and state-of-the-art SSL methods. On CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet, models are first pretrained for 1,000, 1,000, 800, and 100 epochs, respectively, and then evaluated using linear probing. Backbones are highlighted. All reported results are averages obtained from 5 runs on each dataset for each method (including baselines)."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.221,
+ 0.219,
+ 0.789,
+ 0.357
+ ],
+ "angle": 0,
+ "content": "| Method | Loss function | Regularization |
| SimCLR, MoCo | ℓ†NT-Xent = ∑Cc=1C(-1/τic+ sim(zic, zi+c)+ 1/τic- max n=1,...,N sim(zi, zi-c)+ βΩ(τic+)-βΩ(τic-)) | (2.1) |
| SimSiam | ℓ†SymNegCos = ∑Cc=1C(-1/2τic+ sim(zi, i+|sg) - 1/2τic+ sim(zi, i+|h, i|sg) + βΩ(τic+)+ βΩ(τic+)) | (2.2) |
| Barlow Twins | ℓ†Cross-Corr = ∑Cc=1C(d' (1-1/τic+ Cll)2+ λ∑l=1d'∑m≠l d' 1/τic- C2lm + β∑l=1d' 1/τic- d' Ω(τic-)-β∑l=1d' 1/τic- d' Ω(τic-)) | (2.3) |
| LGP, CAN | ℓ†InfoNCE = ∑Cc=1C(-1/τic+ sim(zi, ic+)+ 1/τic- max n=1,...,N+1 sim(zi, ic+±) + βΩ(τic+)-βΩ(τic+)) | (2.4) |
+
+of our AMCL loss for various contrastive learning frameworks. The $\ddagger$ in Eq. (2.1)-(2.4) indicates that these losses represent our multi-head loss versions.
+
+In Eq. (2.1), the asterisk $*$ in $\tau_{in^*}^{c-}$ indicates the index $n^* = \arg \max_{n=1,\dots,N} \sin(\boldsymbol{z}_i, \boldsymbol{z}_{in}^-)$ .
+
+In Eq. (2.2), $\tau_{i}^{c+} = \sigma\big(\langle \phi(\boldsymbol{z}_{i}^{c}), \phi([\boldsymbol{h}_{i}^{+}]_{\mathrm{sg}}) \rangle\big)$ and $\tau_{i}^{c+} = \sigma\big(\langle \phi(\boldsymbol{z}_{i}^{c+}), \phi([\boldsymbol{h}_{i}]_{\mathrm{sg}}) \rangle\big)$ . In Eq. (2.3), temperatures are formed as $\tau_{l}^{c+} = \sigma\big(\langle \phi(\boldsymbol{z}_{l:}^{c}), \phi(\boldsymbol{z}_{l:}^{c+}) \rangle\big)$ and $\tau_{lm}^{c-} = \sigma\big(\langle \phi(\boldsymbol{z}_{l:}^{c}), \phi(\boldsymbol{z}_{m:}^{c+}) \rangle\big)$ where $l \neq m$ and operator ':' simply indexes and concatenates variable as $\boldsymbol{z}_{l} = [z_{l1}, \dots, z_{lN}]^{T}$ . In Eq. (2.4), $\boldsymbol{z}_{in}^{c\pm} = \boldsymbol{z}_{in}^{c-}$ and $\tau_{in}^{c\pm} = \tau_{in}^{c-}$ for $n = 1, \dots, N$ , and $\boldsymbol{z}_{in}^{c\pm} = \boldsymbol{z}_{in}^{c+}$ and $\tau_{in}^{c\pm} = \tau_{i}^{c+}$ if $n = N + 1$ . Finally, notice that for NT-Xent in Eq. (2.1), we have:
+
+$$
+\begin{array}{l} \frac {1}{\tau_ {i n ^ {*}} ^ {c -}} \max _ {n = 1, \dots , N} \sin (\pmb {z} _ {i} ^ {c}, \pmb {z} _ {i n} ^ {c -}) - \varOmega (\tau_ {i n ^ {*}} ^ {c -}) - (2 \pi) ^ {d ^ {\prime} / 2} \\ \approx \log \sum_ {n = 1} ^ {N} \frac {1}{(2 \pi) ^ {d ^ {\prime} / 2} \left(\tau_ {i n} ^ {c -}\right) ^ {d ^ {\prime} / 2}} \exp \left(\frac {1}{\tau_ {i n} ^ {c -}} \left(\operatorname {s i m} \left(z _ {i} ^ {c}, z _ {i n} ^ {c -}\right) - 1\right)\right). \tag {5} \\ \end{array}
+$$
+
+The same approximation (with $z_{in}^{c\pm}$ in place of $z_{in}^{c-}$ ) holds for InfoNCE in Eq. (2.4). Using maximum in Eq. (5), Eq. (2.1) and Eq. (2.4) are somewhat restrictive as the soft-maximum, depending on the temperature, will return the maximum similarity or interpolation over top few similarities close to maximum. Thus, softmax will tackle a group of negative samples closest to the anchor. Indeed, we could use the above soft-maximum in place of maximum but then we have no easy way of recovering the temperature $\tau_{in*}^{c-}$ . Thus, in our experiments we observed that the best choice is to apply $\sum_{k=1}^{\kappa} \frac{1}{\kappa \tau_{in_k}^{c-}} \left[ \operatorname{TopMax}_\kappa \sin(\pmb{z}_i^c, \pmb{z}_{in}^{c-}) \right]_k$ which is the average over top- $\kappa$ largest similarities. As the top- $\kappa$ maximum operation returns also corresponding indexes $n_1^*, \dots, n_\kappa^*$ , the temperature regularization can be easily computed as $\Omega(\{n_k^*\}_{k=1}^\kappa) = \log(\tau_{in_1^*}^{c-} \dots \tau_{in_\kappa^*}^{c-}) + \sum_{k=1}^{\kappa} \frac{1}{\tau_{in_k}^{c-}}$ . The derivation of our loss function is presented in Appendix A.
+
+# 4.4 Discussion
+
+Multiple projection heads vs. a wider MLP projection head. Multiple projection heads offer a significant advantage by creating different sets of
+
+
+(a) head sim. (baseline)
+
+
+(b) head sim. (ours)
+Fig. 3: Distribution of similarity scores for positive and negative pairs. The baseline uses one projection head and constant temperature, while our method has multiple projection heads and adaptive temperature. We use SimCLR for pre-training with ResNet-18 on STL-10. After pre-training, we choose 500 positive pairs and 500 negative pairs from the validation to compute the cosine similarity. In (a) and (b), similarity score (temperature scaled) is computed between the 128-dim features extracted from the projection head(s). In (c) and (d), cosine similarity score is computed between the 512-dim features extracted from the backbone after removing the project heads.
+
+
+(c) feat. sim. (baseline)
+
+
+(d) feat. sim. (ours)
+
+features for each pair of images [43]. This helps build a varied representation that captures various aspects of the input, making the model more resilient and adaptable to different transformations. Furthermore, each projection head's ability to specialize in learning specific patterns or features becomes especially useful when dealing with diverse augmentations that require focused attention on distinct content within the images. In the context of contrastive learning, where diverse augmentation strategies and varying intra-sample similarity introduce more variability in positive pairs, using multiple projection heads excels. This approach effectively manages the resulting variability by enabling the learning of diverse representations. As a result, the model becomes skilled at navigating the intricacies of the data, demonstrating its adaptability to the complex nature of the input. On the other hand, the wider MLP projection head takes a distinct approach by consolidating information from various aspects of the input into a single representation. This proves advantageous when a thorough understanding of the data is crucial, and the relationships between features play a vital role [9].
+
+Visualization of pair similarity distributions. In Fig. 3 we draw the similarity distributions of negative pairs and positive pairs, under the baseline (1 head + constant temperature) and our method (multiple heads + adaptive temperature). When we use the average similarity across the output from the multiple heads, shown in Fig. 3(a) and (b), we can clearly observe better separability brought by our method. It indicates that our method allows for more effective similarity learning of the positive and negative pairs. On the other hand, if we compute the cosine similarity between features extracted right after the backbone, shown in Fig. 3(c) and (d), better separability can again be observed. It illustrates that better similarity learning further benefits representation learning, finally leading to improved linear probing performance.
+
+Connecting our loss function to existing contrastive learning methods. Our loss function consists of three terms: (i) positive temperature-weighted similarities for positive pairs (ii) negative temperature-weighted similarities for negative pairs, and (iii) a regularization term for positive and negative temperatures. As identified in previous works [44, 47], the alignment (closeness) of features
+
+from positive pairs and the uniformity of the induced distribution of the (normalized) features on the hypersphere are the two key properties in contrastive loss. Our loss function also optimizes these properties and further improves the contrastive learning performance through parameterized pair-wise temperature via re-weighting the positive and negative similarities. Our loss function is a more general form, and when we set the temperature to be a global constant, the constant regularization term no longer affects optimization, and thus the loss function reduces to the traditional contrastive loss.
+
+Physical meaning of the regularization term $\Omega(\cdot)$ in Eq. (3). The regularization term consists of regularizing both the positive and negative temperatures. During optimisation, the log $\tau$ term encourages lower positive temperatures and higher negative temperatures, whereas the reverse function term $\frac{1}{\tau}$ is in favour of higher positive temperatures and lower negative temperatures. Hence this regularization term balances the learning of positive and negative temperatures. As we jointly optimize network parameters and the temperature via MLE, this problem naturally becomes decomposed in the maximization of similarities for positive pairs (minimization for negative pairs) weighted by the temperature. However, if $\tau$ was to reach 0 for positive pairs, one would attain a trivial solution. $\Omega(\cdot)$ prevents that trivial solution. Intuitively, one can be very certain in similarity of sample pair but there is a price to pay for that certainty, imposed by $\Omega(\cdot)$ resulting from the Welsch function. This $\Omega(\cdot)$ expresses the prior belief or preference on temperatures. We provide its mathematical insights including connecting temperature to uncertainty in Appendix A. More discussions can be found in Appendix B
+
+# 5 Experiments
+
+We choose popular datasets that are widely used in evaluating the SSL models, including CIFAR-10, CIFAR-100, STL-10, Tiny-[ImageNet and ImageNet. The dataset details are provided in Appendix Below, we describe the experimental setup and evaluations.
+
+# 5.1 Setup
+
+We conduct experiments on the aforementioned datasets following practices outlined by [25]. We consider five different types of transformations for data augmentations: random cropping, random Gaussian blur, color dropping (i.e., randomly converting images to grayscale), color distortion, and random horizontal flipping. For pre-training, we employ Resnet-18 (R18) [22] for SimCLR [9], MoCo [21], SimSiam [10], and Barlow Twins [51] on CIFAR-10, CIFAR-100 and STL-10. For Tiny-ImageNet and ImageNet, we use Resnet-50 (R50) variants, ViT-B and ViT-L. Other settings, such as the architecture of the projection head, remain consistent with the original algorithm configurations. We select $C$ projection heads in range from 2 to 6. We set the temperature bounds ( $\eta$ and $\iota$ of sigmoid) in range $[1e - 5, 2]$ on smaller datasets and $[1e - 5, 5]$ for ImageNet and
+
+Table 3: Impact of AMCL when applied to popular and state-of-the-art SSL methods. On CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet, models are first pretrained for 1,000, 1,000, 800, and 100 epochs, respectively, and then evaluated using linear probing. Backbones are highlighted. All reported results are averages obtained from 5 runs on each dataset for each method (including baselines).
+
+| Method | Loss function | Regularization |
| SimCLR, MoCo | ℓ†NT-Xent = ∑Cc=1C(-1/τic+ sim(zic, zi+c)+ 1/τic- max n=1,...,N sim(zi, zi-c)+ βΩ(τic+)-βΩ(τic-)) | (2.1) |
| SimSiam | ℓ†SymNegCos = ∑Cc=1C(-1/2τic+ sim(zi, i+|sg) - 1/2τic+ sim(zi, i+|h, i|sg) + βΩ(τic+)+ βΩ(τic+)) | (2.2) |
| Barlow Twins | ℓ†Cross-Corr = ∑Cc=1C(d' (1-1/τic+ Cll)2+ λ∑l=1d'∑m≠l d' 1/τic- C2lm + β∑l=1d' 1/τic- d' Ω(τic-)-β∑l=1d' 1/τic- d' Ω(τic-)) | (2.3) |
| LGP, CAN | ℓ†InfoNCE = ∑Cc=1C(-1/τic+ sim(zi, ic+)+ 1/τic- max n=1,...,N+1 sim(zi, ic+±) + βΩ(τic+)-βΩ(τic+)) | (2.4) |
",
+ "image_path": "966624a75b897b45f70dffa6d37ffda8e9d4574fc39ef59373c2faaa11304e76.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 130,
+ 244,
+ 479,
+ 268
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 244,
+ 479,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 244,
+ 479,
+ 268
+ ],
+ "type": "text",
+ "content": "of our AMCL loss for various contrastive learning frameworks. The "
+ },
+ {
+ "bbox": [
+ 130,
+ 244,
+ 479,
+ 268
+ ],
+ "type": "inline_equation",
+ "content": "\\ddagger"
+ },
+ {
+ "bbox": [
+ 130,
+ 244,
+ 479,
+ 268
+ ],
+ "type": "text",
+ "content": " in Eq. (2.1)-(2.4) indicates that these losses represent our multi-head loss versions."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "text",
+ "content": "In Eq. (2.1), the asterisk "
+ },
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "inline_equation",
+ "content": "*"
+ },
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "text",
+ "content": " in "
+ },
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{in^*}^{c-}"
+ },
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "text",
+ "content": " indicates the index "
+ },
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "inline_equation",
+ "content": "n^* = \\arg \\max_{n=1,\\dots,N} \\sin(\\boldsymbol{z}_i, \\boldsymbol{z}_{in}^-)"
+ },
+ {
+ "bbox": [
+ 148,
+ 269,
+ 490,
+ 289
+ ],
+ "type": "text",
+ "content": "."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": "In Eq. (2.2), "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{i}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{i}^{c}), \\phi([\\boldsymbol{h}_{i}^{+}]_{\\mathrm{sg}}) \\rangle\\big)"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{i}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{i}^{c+}), \\phi([\\boldsymbol{h}_{i}]_{\\mathrm{sg}}) \\rangle\\big)"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": ". In Eq. (2.3), temperatures are formed as "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{l}^{c+} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{l:}^{c}), \\phi(\\boldsymbol{z}_{l:}^{c+}) \\rangle\\big)"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{lm}^{c-} = \\sigma\\big(\\langle \\phi(\\boldsymbol{z}_{l:}^{c}), \\phi(\\boldsymbol{z}_{m:}^{c+}) \\rangle\\big)"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " where "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "l \\neq m"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " and operator ':' simply indexes and concatenates variable as "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\boldsymbol{z}_{l} = [z_{l1}, \\dots, z_{lN}]^{T}"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": ". In Eq. (2.4), "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\boldsymbol{z}_{in}^{c\\pm} = \\boldsymbol{z}_{in}^{c-}"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{in}^{c\\pm} = \\tau_{in}^{c-}"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " for "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "n = 1, \\dots, N"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": ", and "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\boldsymbol{z}_{in}^{c\\pm} = \\boldsymbol{z}_{in}^{c+}"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{in}^{c\\pm} = \\tau_{i}^{c+}"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": " if "
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "inline_equation",
+ "content": "n = N + 1"
+ },
+ {
+ "bbox": [
+ 130,
+ 289,
+ 481,
+ 365
+ ],
+ "type": "text",
+ "content": ". Finally, notice that for NT-Xent in Eq. (2.1), we have:"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 183,
+ 375,
+ 481,
+ 436
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 183,
+ 375,
+ 481,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 183,
+ 375,
+ 481,
+ 436
+ ],
+ "type": "interline_equation",
+ "content": "\\begin{array}{l} \\frac {1}{\\tau_ {i n ^ {*}} ^ {c -}} \\max _ {n = 1, \\dots , N} \\sin (\\pmb {z} _ {i} ^ {c}, \\pmb {z} _ {i n} ^ {c -}) - \\varOmega (\\tau_ {i n ^ {*}} ^ {c -}) - (2 \\pi) ^ {d ^ {\\prime} / 2} \\\\ \\approx \\log \\sum_ {n = 1} ^ {N} \\frac {1}{(2 \\pi) ^ {d ^ {\\prime} / 2} \\left(\\tau_ {i n} ^ {c -}\\right) ^ {d ^ {\\prime} / 2}} \\exp \\left(\\frac {1}{\\tau_ {i n} ^ {c -}} \\left(\\operatorname {s i m} \\left(z _ {i} ^ {c}, z _ {i n} ^ {c -}\\right) - 1\\right)\\right). \\tag {5} \\\\ \\end{array}",
+ "image_path": "3014dce9fe37fda24a2aa954e1bf96336c26001725de7a26492fbe2f6f5f0aa8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": "The same approximation (with "
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "z_{in}^{c\\pm}"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": " in place of "
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "z_{in}^{c-}"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": ") holds for InfoNCE in Eq. (2.4). Using maximum in Eq. (5), Eq. (2.1) and Eq. (2.4) are somewhat restrictive as the soft-maximum, depending on the temperature, will return the maximum similarity or interpolation over top few similarities close to maximum. Thus, softmax will tackle a group of negative samples closest to the anchor. Indeed, we could use the above soft-maximum in place of maximum but then we have no easy way of recovering the temperature "
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "\\tau_{in*}^{c-}"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": ". Thus, in our experiments we observed that the best choice is to apply "
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "\\sum_{k=1}^{\\kappa} \\frac{1}{\\kappa \\tau_{in_k}^{c-}} \\left[ \\operatorname{TopMax}_\\kappa \\sin(\\pmb{z}_i^c, \\pmb{z}_{in}^{c-}) \\right]_k"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": " which is the average over top-"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "\\kappa"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": " largest similarities. As the top-"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "\\kappa"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": " maximum operation returns also corresponding indexes "
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "n_1^*, \\dots, n_\\kappa^*"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": ", the temperature regularization can be easily computed as "
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "inline_equation",
+ "content": "\\Omega(\\{n_k^*\\}_{k=1}^\\kappa) = \\log(\\tau_{in_1^*}^{c-} \\dots \\tau_{in_\\kappa^*}^{c-}) + \\sum_{k=1}^{\\kappa} \\frac{1}{\\tau_{in_k}^{c-}}"
+ },
+ {
+ "bbox": [
+ 130,
+ 445,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": ". The derivation of our loss function is presented in Appendix A."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 132,
+ 620,
+ 214,
+ 632
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 620,
+ 214,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 620,
+ 214,
+ 632
+ ],
+ "type": "text",
+ "content": "4.4 Discussion"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 130,
+ 641,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 641,
+ 482,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 641,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "content": "Multiple projection heads vs. a wider MLP projection head. Multiple projection heads offer a significant advantage by creating different sets of"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 273,
+ 91,
+ 448,
+ 102
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 273,
+ 91,
+ 448,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 273,
+ 91,
+ 448,
+ 102
+ ],
+ "type": "text",
+ "content": "Adaptive Multi-head Contrastive Learning"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 474,
+ 92,
+ 480,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 480,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 480,
+ 100
+ ],
+ "type": "text",
+ "content": "7"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 133,
+ 114,
+ 220,
+ 175
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 220,
+ 175
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 220,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 220,
+ 175
+ ],
+ "type": "image",
+ "image_path": "6079c683a06bc7c366d37dda567ba0a16a2d58164fb00e7442cba050bbdcad93.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 134,
+ 176,
+ 219,
+ 185
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 134,
+ 176,
+ 219,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 134,
+ 176,
+ 219,
+ 185
+ ],
+ "type": "text",
+ "content": "(a) head sim. (baseline)"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 223,
+ 114,
+ 306,
+ 175
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 223,
+ 114,
+ 306,
+ 175
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 223,
+ 114,
+ 306,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 223,
+ 114,
+ 306,
+ 175
+ ],
+ "type": "image",
+ "image_path": "3b8f6cdf33cb7fe9f4e8c9895538d18a154551376b6608265dbd13c474460bb0.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 227,
+ 176,
+ 299,
+ 186
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 227,
+ 176,
+ 299,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 227,
+ 176,
+ 299,
+ 186
+ ],
+ "type": "text",
+ "content": "(b) head sim. (ours)"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 130,
+ 189,
+ 482,
+ 277
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 189,
+ 482,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 189,
+ 482,
+ 277
+ ],
+ "type": "text",
+ "content": "Fig. 3: Distribution of similarity scores for positive and negative pairs. The baseline uses one projection head and constant temperature, while our method has multiple projection heads and adaptive temperature. We use SimCLR for pre-training with ResNet-18 on STL-10. After pre-training, we choose 500 positive pairs and 500 negative pairs from the validation to compute the cosine similarity. In (a) and (b), similarity score (temperature scaled) is computed between the 128-dim features extracted from the projection head(s). In (c) and (d), cosine similarity score is computed between the 512-dim features extracted from the backbone after removing the project heads."
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 307,
+ 114,
+ 391,
+ 175
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 114,
+ 391,
+ 175
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 114,
+ 391,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 114,
+ 391,
+ 175
+ ],
+ "type": "image",
+ "image_path": "935b4b8f60ce8c4e01b1e2d5cebb32fbbeb7af7177988b0b60e8b8c98dc30fb3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 307,
+ 176,
+ 390,
+ 186
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 176,
+ 390,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 176,
+ 390,
+ 186
+ ],
+ "type": "text",
+ "content": "(c) feat. sim. (baseline)"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 391,
+ 114,
+ 477,
+ 175
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 391,
+ 114,
+ 477,
+ 175
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 391,
+ 114,
+ 477,
+ 175
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 391,
+ 114,
+ 477,
+ 175
+ ],
+ "type": "image",
+ "image_path": "26e08683f10278536ebf85eaca98a8368a749a9babfc9e34f4d32a3c0a53a66c.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 398,
+ 176,
+ 469,
+ 186
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 398,
+ 176,
+ 469,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 398,
+ 176,
+ 469,
+ 186
+ ],
+ "type": "text",
+ "content": "(d) feat. sim. (ours)"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 130,
+ 290,
+ 482,
+ 471
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 290,
+ 482,
+ 471
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 290,
+ 482,
+ 471
+ ],
+ "type": "text",
+ "content": "features for each pair of images [43]. This helps build a varied representation that captures various aspects of the input, making the model more resilient and adaptable to different transformations. Furthermore, each projection head's ability to specialize in learning specific patterns or features becomes especially useful when dealing with diverse augmentations that require focused attention on distinct content within the images. In the context of contrastive learning, where diverse augmentation strategies and varying intra-sample similarity introduce more variability in positive pairs, using multiple projection heads excels. This approach effectively manages the resulting variability by enabling the learning of diverse representations. As a result, the model becomes skilled at navigating the intricacies of the data, demonstrating its adaptability to the complex nature of the input. On the other hand, the wider MLP projection head takes a distinct approach by consolidating information from various aspects of the input into a single representation. This proves advantageous when a thorough understanding of the data is crucial, and the relationships between features play a vital role [9]."
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 130,
+ 472,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 472,
+ 482,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 472,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": "Visualization of pair similarity distributions. In Fig. 3 we draw the similarity distributions of negative pairs and positive pairs, under the baseline (1 head + constant temperature) and our method (multiple heads + adaptive temperature). When we use the average similarity across the output from the multiple heads, shown in Fig. 3(a) and (b), we can clearly observe better separability brought by our method. It indicates that our method allows for more effective similarity learning of the positive and negative pairs. On the other hand, if we compute the cosine similarity between features extracted right after the backbone, shown in Fig. 3(c) and (d), better separability can again be observed. It illustrates that better similarity learning further benefits representation learning, finally leading to improved linear probing performance."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 130,
+ 605,
+ 482,
+ 667
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 605,
+ 482,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 605,
+ 482,
+ 667
+ ],
+ "type": "text",
+ "content": "Connecting our loss function to existing contrastive learning methods. Our loss function consists of three terms: (i) positive temperature-weighted similarities for positive pairs (ii) negative temperature-weighted similarities for negative pairs, and (iii) a regularization term for positive and negative temperatures. As identified in previous works [44, 47], the alignment (closeness) of features"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "text",
+ "content": "8"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 166,
+ 90,
+ 425,
+ 103
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 90,
+ 425,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 90,
+ 425,
+ 103
+ ],
+ "type": "text",
+ "content": "Lei Wang, Piotr Koniusz, Tom Gedeon, and Liang Zheng"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 7
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 482,
+ 209
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 482,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 482,
+ 209
+ ],
+ "type": "text",
+ "content": "from positive pairs and the uniformity of the induced distribution of the (normalized) features on the hypersphere are the two key properties in contrastive loss. Our loss function also optimizes these properties and further improves the contrastive learning performance through parameterized pair-wise temperature via re-weighting the positive and negative similarities. Our loss function is a more general form, and when we set the temperature to be a global constant, the constant regularization term no longer affects optimization, and thus the loss function reduces to the traditional contrastive loss."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": "Physical meaning of the regularization term "
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "inline_equation",
+ "content": "\\Omega(\\cdot)"
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": " in Eq. (3). The regularization term consists of regularizing both the positive and negative temperatures. During optimisation, the log "
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "inline_equation",
+ "content": "\\tau"
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": " term encourages lower positive temperatures and higher negative temperatures, whereas the reverse function term "
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "inline_equation",
+ "content": "\\frac{1}{\\tau}"
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": " is in favour of higher positive temperatures and lower negative temperatures. Hence this regularization term balances the learning of positive and negative temperatures. As we jointly optimize network parameters and the temperature via MLE, this problem naturally becomes decomposed in the maximization of similarities for positive pairs (minimization for negative pairs) weighted by the temperature. However, if "
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "inline_equation",
+ "content": "\\tau"
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": " was to reach 0 for positive pairs, one would attain a trivial solution. "
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "inline_equation",
+ "content": "\\Omega(\\cdot)"
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": " prevents that trivial solution. Intuitively, one can be very certain in similarity of sample pair but there is a price to pay for that certainty, imposed by "
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "inline_equation",
+ "content": "\\Omega(\\cdot)"
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": " resulting from the Welsch function. This "
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "inline_equation",
+ "content": "\\Omega(\\cdot)"
+ },
+ {
+ "bbox": [
+ 130,
+ 211,
+ 483,
+ 404
+ ],
+ "type": "text",
+ "content": " expresses the prior belief or preference on temperatures. We provide its mathematical insights including connecting temperature to uncertainty in Appendix A. More discussions can be found in Appendix B"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 132,
+ 422,
+ 230,
+ 436
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 422,
+ 230,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 422,
+ 230,
+ 436
+ ],
+ "type": "text",
+ "content": "5 Experiments"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 130,
+ 447,
+ 482,
+ 495
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 447,
+ 482,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 447,
+ 482,
+ 495
+ ],
+ "type": "text",
+ "content": "We choose popular datasets that are widely used in evaluating the SSL models, including CIFAR-10, CIFAR-100, STL-10, Tiny-[ImageNet and ImageNet. The dataset details are provided in Appendix Below, we describe the experimental setup and evaluations."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 132,
+ 514,
+ 190,
+ 526
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 514,
+ 190,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 514,
+ 190,
+ 526
+ ],
+ "type": "text",
+ "content": "5.1 Setup"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "text",
+ "content": "We conduct experiments on the aforementioned datasets following practices outlined by [25]. We consider five different types of transformations for data augmentations: random cropping, random Gaussian blur, color dropping (i.e., randomly converting images to grayscale), color distortion, and random horizontal flipping. For pre-training, we employ Resnet-18 (R18) [22] for SimCLR [9], MoCo [21], SimSiam [10], and Barlow Twins [51] on CIFAR-10, CIFAR-100 and STL-10. For Tiny-ImageNet and ImageNet, we use Resnet-50 (R50) variants, ViT-B and ViT-L. Other settings, such as the architecture of the projection head, remain consistent with the original algorithm configurations. We select "
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "inline_equation",
+ "content": "C"
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "text",
+ "content": " projection heads in range from 2 to 6. We set the temperature bounds ("
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "inline_equation",
+ "content": "\\eta"
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "text",
+ "content": " and "
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "inline_equation",
+ "content": "\\iota"
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "text",
+ "content": " of sigmoid) in range "
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "inline_equation",
+ "content": "[1e - 5, 2]"
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "text",
+ "content": " on smaller datasets and "
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "inline_equation",
+ "content": "[1e - 5, 5]"
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 668
+ ],
+ "type": "text",
+ "content": " for ImageNet and"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 273,
+ 91,
+ 448,
+ 102
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 273,
+ 91,
+ 448,
+ 102
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 273,
+ 91,
+ 448,
+ 102
+ ],
+ "type": "text",
+ "content": "Adaptive Multi-head Contrastive Learning"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 474,
+ 92,
+ 481,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 481,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 481,
+ 100
+ ],
+ "type": "text",
+ "content": "9"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 8
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 135,
+ 173,
+ 482,
+ 282
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 130,
+ 114,
+ 482,
+ 170
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 114,
+ 482,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 114,
+ 482,
+ 170
+ ],
+ "type": "text",
+ "content": "Table 3: Impact of AMCL when applied to popular and state-of-the-art SSL methods. On CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet, models are first pretrained for 1,000, 1,000, 800, and 100 epochs, respectively, and then evaluated using linear probing. Backbones are highlighted. All reported results are averages obtained from 5 runs on each dataset for each method (including baselines)."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 135,
+ 173,
+ 482,
+ 282
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 135,
+ 173,
+ 482,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 135,
+ 173,
+ 482,
+ 282
+ ],
+ "type": "table",
+ "html": "| Visual input examples for Image Address Localization: |
| Grant Street, Downtown | Fifth Avenue, Downtown |
| User | Where might this photo have been taken? Tell me its street level address. |
| LLaVA-IAL | The address of this photo might be Grant Street, Downtown, Pittsburgh, PA, USA. | The address of this photo might be Fifth Avenue, Downtown, Pittsburgh, PA, USA. |
| GPT-4V [1] | This photo was taken in Peavey Plaza in Minneapolis, Minnesota, USA, on the corner of Nicolson Mall and 11th Street. | This photo was taken at 400 Fifth Avenue in Pittsburgh, Pennsylvania. |
| Gemini-Pro-V [44] | This photo was taken on the Rose Fitzgerald Kennedy Greenway in Boston, Massachusetts. | The photo was taken on Fifth Avenue in Pittsburgh, Pennsylvania. |
| QWen-VL [5] | This photo appears to have been taken on a city street, possibly in a downtown area or commercial district. | This photo appears to have been taken on Fifth Avenue in Pittsburgh, Pennsylvania, USA. The street sign in the image confirms this location. |
",
+ "bbox": [
+ 220,
+ 176,
+ 772,
+ 392
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "6.6 Limitations and Future Work",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 422,
+ 504,
+ 436
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "The proposed AddressCLIP can be regarded as a discriminative model, limited by the set of candidate addresses at inference. In contrast, generative models such as multimodal large language models (MLLMs) [1, 5, 44] can yield more flexible and interactive geographic textual information but may not offer precise administrative addresses. In future work, we plan to explore the potential of MLLMs for the IAL task. To this end, we have made an attempt by constructing a question-and-answer dataset based on Pitts-IAL and adopting the LLaVA-1.5 [31] model for instruct tuning. Tab. 7 shows some examples comparing with representative MLLMs. It is evident that LLaVA-IAL, fine-tuned with instructions, can predict accurate administrative addresses consistent with address hierarchy, while other compared MLLMs are unable to predict addresses without landmarks or street signs and lack standardized output formats.",
+ "bbox": [
+ 212,
+ 449,
+ 787,
+ 630
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "7 Conclusion",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 655,
+ 359,
+ 671
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "In this study, we introduce the problem of image address localization and propose three IAL datasets for evaluation and subsequent research. To facilitate the alignment of images and addresses for tackling the problem, we propose the AddressCLIP framework consisting of image-text alignment and image-geography matching. Extensive experiments on the proposed datasets validate that our method outperforms transfer learning methods that transfer CLIP to downstream tasks. We compare the proposed method with the existing two-stage address localization pipeline based on the image geo-localization technology and discuss AddressCLIP's application in real-world situations. Finally, we explore the potential of multimodal large language models for address localization.",
+ "bbox": [
+ 212,
+ 688,
+ 787,
+ 840
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "page_number",
+ "text": "14",
+ "bbox": [
+ 217,
+ 114,
+ 235,
+ 126
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "header",
+ "text": "S. Xu et al.",
+ "bbox": [
+ 271,
+ 114,
+ 349,
+ 127
+ ],
+ "page_idx": 13
+ },
+ {
+ "type": "text",
+ "text": "Acknowledgements",
+ "text_level": 1,
+ "bbox": [
+ 217,
+ 143,
+ 401,
+ 162
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "This work was supported by the National Natural Science Foundations of China (Grants No.62376267, 62076242) and the innoHK project.",
+ "bbox": [
+ 215,
+ 178,
+ 784,
+ 209
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "text",
+ "text": "References",
+ "text_level": 1,
+ "bbox": [
+ 217,
+ 233,
+ 321,
+ 250
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)",
+ "2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022)",
+ "3. Ali-Bey, A., Chaib-Draa, B., Giguere, P.: Mixvpr: Feature mixing for visual place recognition. In: WACV. pp. 2998-3007 (2023)",
+ "4. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: CVPR. pp. 5297-5307 (2016)",
+ "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)",
+ "6. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer vision and image understanding 110(3), 346-359 (2008)",
+ "7. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: CVPR. pp. 4878-4888 (2022)",
+ "8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023)",
+ "9. Berton, G.M., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: WACV. pp. 2918-2927 (2021)",
+ "0. Cepeda, V.V., Nayak, G.K., Shah, M.: Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. arXiv preprint arXiv:2309.16020 (2023)",
+ "1. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)",
+ "2. Clark, B., Kerrigan, A., Kulkarni, P.P., Cepeda, V.V., Shah, M.: Where we are and what we're looking at: Query based worldwide image geo-localization using hierarchies and scenes. In: CVPR. pp. 23182-23190 (2023)",
+ "3. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: ECCV Workshop. vol. 1, pp. 1-2 (2004)",
+ "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)",
+ "5. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)",
+ "6. Ge, Y., Wang, H., Zhu, F., Zhao, R., Li, H.: Self-supervising fine-grained region similarities for large-scale image localization. In: ECCV. pp. 369-386 (2020)"
+ ],
+ "bbox": [
+ 225,
+ 266,
+ 785,
+ 839
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "header",
+ "text": "AddressCLIP",
+ "bbox": [
+ 638,
+ 114,
+ 730,
+ 126
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "page_number",
+ "text": "15",
+ "bbox": [
+ 767,
+ 116,
+ 785,
+ 126
+ ],
+ "page_idx": 14
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "17. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: End-to-end learning of deep visual representations for image retrieval. International Journal of Computer Vision 124(2), 237-254 (2017)",
+ "18. Haas, L., Alberti, S., Skreta, M.: Learning generalized zero-shot learners for open-domain image geolocation. arXiv preprint arXiv:2302.00275 (2023)",
+ "19. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: CVPR. pp. 14141-14152 (2021)",
+ "20. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020)",
+ "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)",
+ "22. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)",
+ "23. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)",
+ "24. Jegou, H., Douze, M., Schmid, C.: Hamming embedding and weak geometric consistency for large scale image search. In: ECCV. pp. 304-317 (2008)",
+ "25. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE transactions on pattern analysis and machine intelligence 34(9), 1704-1716 (2011)",
+ "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)",
+ "27. Jin Kim, H., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geo-localization. In: CVPR. pp. 2136-2145 (2017)",
+ "28. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)",
+ "29. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR. pp. 19113-19122 (2023)",
+ "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900 (2022)",
+ "31. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)",
+ "32. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: ICCV. pp. 2570-2579 (2019)",
+ "33. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)",
+ "34. Oliva, A., Torralba, A.: Building the gist of a scene: The role of global image features in recognition. Progress in brain research 155, 23-36 (2006)",
+ "35. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)",
+ "36. Pramanick, S., Nowara, E.M., Gleason, J., Castillo, C.D., Chellappa, R.: Where in the world is this image? transformer-based geo-localization in the wild. In: ECCV. pp. 196-215 (2022)",
+ "37. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE transactions on pattern analysis and machine intelligence 41(7), 1655-1668 (2018)"
+ ],
+ "bbox": [
+ 215,
+ 147,
+ 784,
+ 839
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "page_number",
+ "text": "16",
+ "bbox": [
+ 217,
+ 114,
+ 235,
+ 126
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "header",
+ "text": "S. Xu et al.",
+ "bbox": [
+ 271,
+ 114,
+ 349,
+ 127
+ ],
+ "page_idx": 15
+ },
+ {
+ "type": "list",
+ "sub_type": "ref_text",
+ "list_items": [
+ "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)",
+ "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)",
+ "40. Schindler, G., Brown, M., Szeliski, R.: City-scale location recognition. In: CVPR. pp. 1-7. IEEE (2007)",
+ "41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv abs/2210.08402 (2022), https://api_semanticscholar.org/CorpusID: 252917726",
+ "42. Seo, P.H., Weyand, T., Sim, J., Han, B.: Cplanet: Enhancing image geolocation by combinatorial partitioning of maps. In: ECCV. pp. 536-551 (2018)",
+ "43. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)",
+ "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)",
+ "45. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: CVPR. pp. 883-890 (2013)",
+ "46. Trivigno, G., Berton, G., Aragon, J., Caputo, B., Masone, C.: Divide&classify: Fine-grained classification for city-wide visual geo-localization. In: ICCV. pp. 11142-11152 (2023)",
+ "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)",
+ "48. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: CVPR. pp. 13648-13657 (2022)",
+ "49. Weyand, T., Kostrikov, I., Philbin, J.: Planet-photo geolocation with convolutional neural networks. In: ECCV. pp. 37-55 (2016)",
+ "50. Wilson, D., Zhang, X., Sultani, W., Wshah, S.: Image and object geo-localization. International Journal of Computer Vision pp. 1-43 (2023)",
+ "51. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022)",
+ "52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)",
+ "53. Zhu, S., Yang, L., Chen, C., Shah, M., Shen, X., Wang, H.: R2former: Unified retrieval and reranking transformer for place recognition. In: CVPR. pp. 19370-19380 (2023)",
+ "54. Zhu, Y., Wang, J., Xie, L., Zheng, L.: Attention-based pyramid aggregation network for visual place recognition. In: ACM MM. pp. 99-107 (2018)"
+ ],
+ "bbox": [
+ 212,
+ 146,
+ 787,
+ 756
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "header",
+ "text": "AddressCLIP",
+ "bbox": [
+ 638,
+ 114,
+ 730,
+ 126
+ ],
+ "page_idx": 16
+ },
+ {
+ "type": "page_number",
+ "text": "17",
+ "bbox": [
+ 767,
+ 114,
+ 785,
+ 126
+ ],
+ "page_idx": 16
+ }
+]
\ No newline at end of file
diff --git a/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/bbd4751a-6e05-4f3a-882f-4ee98bbd88b8_model.json b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/bbd4751a-6e05-4f3a-882f-4ee98bbd88b8_model.json
new file mode 100644
index 0000000000000000000000000000000000000000..6d361dcf35b61b3b560a618715169cdb93d3587a
--- /dev/null
+++ b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/bbd4751a-6e05-4f3a-882f-4ee98bbd88b8_model.json
@@ -0,0 +1,2687 @@
+[
+ [
+ {
+ "type": "title",
+ "bbox": [
+ 0.219,
+ 0.141,
+ 0.784,
+ 0.187
+ ],
+ "angle": 0,
+ "content": "AddressCLIP: Empowering Vision-Language Models for City-wide Image Address Localization"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.218,
+ 0.212,
+ 0.784,
+ 0.245
+ ],
+ "angle": 0,
+ "content": "Shixiong \\(\\mathrm{Xu}^{1,3*†}\\), Chenghao Zhang\\(^{2*}\\), Lubin Fan\\(^{2‡}\\), Gaofeng Meng\\(^{1,3,4‡}\\), Shiming Xiang\\(^{1,3}\\), and Jieping Ye\\(^{2}\\)"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.287,
+ 0.255,
+ 0.714,
+ 0.282
+ ],
+ "angle": 0,
+ "content": "\\(^{1}\\) MAIS, Institute of Automation, Chinese Academy of Sciences \\(^{2}\\) Alibaba Cloud"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.243,
+ 0.283,
+ 0.757,
+ 0.298
+ ],
+ "angle": 0,
+ "content": "\\(^{3}\\) School of Artificial Intelligence, University of Chinese Academy of Sciences"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.243,
+ 0.297,
+ 0.757,
+ 0.311
+ ],
+ "angle": 0,
+ "content": "\\(^{4}\\) CAIR, HK Institute of Science & Innovation, Chinese Academy of Sciences"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.243,
+ 0.283,
+ 0.757,
+ 0.311
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.261,
+ 0.355,
+ 0.744,
+ 0.606
+ ],
+ "angle": 0,
+ "content": "Abstract. In this study, we introduce a new problem raised by social media and photojournalism, named Image Address Localization (IAL), which aims to predict the readable textual address where an image was taken. Existing two-stage approaches involve predicting geographical coordinates and converting them into human-readable addresses, which can lead to ambiguity and be resource-intensive. In contrast, we propose an end-to-end framework named AddressCLIP to solve the problem with more semantics, consisting of two key ingredients: i) image-text alignment to align images with addresses and scene captions by contrastive learning, and ii) image-geography matching to constrain image features with the spatial distance in terms of manifold learning. Additionally, we have built three datasets from Pittsburgh and San Francisco on different scales specifically for the IAL problem. Experiments demonstrate that our approach achieves compelling performance on the proposed datasets and outperforms representative transfer learning methods for vision-language models. Furthermore, extensive ablations and visualizations exhibit the effectiveness of the proposed method. The datasets and source code are available at https://github.com/xsx1001/AddressCLIP."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.261,
+ 0.618,
+ 0.741,
+ 0.647
+ ],
+ "angle": 0,
+ "content": "Keywords: Image address localization \\(\\cdot\\) Image-text alignment \\(\\cdot\\) Image-geography matching \\(\\cdot\\) Vision-language model"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.217,
+ 0.675,
+ 0.377,
+ 0.691
+ ],
+ "angle": 0,
+ "content": "1 Introduction"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.71,
+ 0.788,
+ 0.787
+ ],
+ "angle": 0,
+ "content": "Users on social media platforms such as Facebook and Instagram often tag their pictures with textual addresses to connect with local communities, raising the demand for predicting the descriptive address information of the place where an image was taken. This has various practical applications, for instance, businesses and travel platforms can use addresses of images to provide recommendations"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.232,
+ 0.797,
+ 0.721,
+ 0.812
+ ],
+ "angle": 0,
+ "content": "This work was done when Shixiong Xu was an intern at Alibaba Cloud."
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.235,
+ 0.812,
+ 0.38,
+ 0.825
+ ],
+ "angle": 0,
+ "content": "* Equal contributions"
+ },
+ {
+ "type": "page_footnote",
+ "bbox": [
+ 0.234,
+ 0.825,
+ 0.399,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "‡ Corresponding authors"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.232,
+ 0.797,
+ 0.721,
+ 0.84
+ ],
+ "angle": 0,
+ "content": null
+ }
+ ],
+ [
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.218,
+ 0.116,
+ 0.23,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "2"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.272,
+ 0.115,
+ 0.351,
+ 0.128
+ ],
+ "angle": 0,
+ "content": "S. Xu et al."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.218,
+ 0.145,
+ 0.482,
+ 0.297
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.491,
+ 0.145,
+ 0.788,
+ 0.297
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.214,
+ 0.303,
+ 0.789,
+ 0.348
+ ],
+ "angle": 0,
+ "content": "Fig. 1: Comparison of image-based geo-localization and address localization tasks. The objective of the proposed task is to predict the semantic text address of a given image instead of a digital GPS coordinate without the need for a retrieval gallery."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.38,
+ 0.784,
+ 0.41
+ ],
+ "angle": 0,
+ "content": "or organize location-specific content. Additionally, photojournalism can rapidly verify the authenticity of the event with the image's address."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.413,
+ 0.789,
+ 0.579
+ ],
+ "angle": 0,
+ "content": "To predict an image's address, one reasonable approach involves leveraging image geo-localization technology to predict GPS coordinates (i.e., latitude and longitude) from an image [50], followed by the reverse Geocoding to query for a readable address. Image geo-localization, also known as visual place recognition, is commonly treated as an image retrieval problem where a database of geo-tagged images serves as a matching reference for the query image. Previous retrieval-based methods [3,4,7,19,28] have shown remarkable performance. However, in practice, the creation of pre-collected geo-tagged databases requires significant labor and storage resources, while GPS coordinates lack readability and semantics. In addition, the conversion from GPS to readable addresses often presents ambiguities, and the Image-GPS-Address pipeline is not end-to-end."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.582,
+ 0.789,
+ 0.763
+ ],
+ "angle": 0,
+ "content": "To alleviate the above issues, in this study, we propose to perform Image Address Localization (IAL) where a model is tasked to predict the readable textual address where a given image was taken. We design a semantic address partition strategy to perform fine-grained partitioning of city-wide addresses, conforming to the way humans describe address information. By doing this, we are able to train models in an end-to-end manner, and during inference, there is no need to construct a retrieval database which greatly reduces the storage and retrieval burden. Furthermore, the model's output addresses align more closely with human description habits, which provides a bridge for subsequent city-wide scene understanding and point-of-interest recommendation. Fig. 1 shows the comparison of image geo-localization and image address localization tasks, where the latter focuses on predicting human-readable textual address information."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.765,
+ 0.789,
+ 0.842
+ ],
+ "angle": 0,
+ "content": "In this study, we propose an end-to-end framework, AddressCLIP, based on the visual-language model CLIP [38], aiming to learn an alignment of images and addresses. Our approach leverages two key ingredients: image-text alignment and image-geography matching. Firstly, we introduce additional scene captions as a supplement to address text thus facilitating the alignment of images and tex"
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.641,
+ 0.115,
+ 0.732,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "AddressCLIP"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.775,
+ 0.117,
+ 0.785,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "3"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.147,
+ 0.784,
+ 0.192
+ ],
+ "angle": 0,
+ "content": "tual addresses by contrastive learning. Secondly, we propose an image-geography matching mechanism to bring features of geographically proximate images closer while separating features of images that are far apart geographically."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.192,
+ 0.787,
+ 0.313
+ ],
+ "angle": 0,
+ "content": "To support the image address localization task, we constructed three IAL datasets of different sizes based on the Pitts-250k [4] and SF-XL datasets [7]: Pitts-IAL (234K), SF-IAL-Base (184K), and SF-IAL-Large (1.96M). In contrast to the original datasets, each image in our dataset is accompanied by not only its geographical coordinate but also the administrative address. Specifically, we utilized the reverse Geocoding API of Google Maps to retrieve administrative addresses for a portion of the images and obtain addresses for the remaining images through nearest-neighbor interpolation of the geographical coordinates."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.313,
+ 0.788,
+ 0.464
+ ],
+ "angle": 0,
+ "content": "We evaluate the proposed AddressCLIP framework on the introduced datasets. Our proposed method achieves a Top-1 address localization accuracy of over \\(80\\%\\) across three IAL datasets, most notably reaching a performance of \\(85.92\\%\\) on the largest dataset, SF-IAL-Large. Compared with challenging baselines [29, 51, 52] that transfer CLIP to the downstream IAL task, our AddressCLIP achieves improvements of \\(3\\%\\) to \\(6\\%\\) on the proposed datasets. In addition, the qualitative results demonstrate good alignment between images and textual address queries in geographical space. Finally, we discuss the superiority of the proposed method over the two-stage \"Image-GPS-Address\" approach and explore the application prospects of multimodal large language models in the IAL task."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.24,
+ 0.465,
+ 0.568,
+ 0.479
+ ],
+ "angle": 0,
+ "content": "Our contributions are summarized as follows:"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.226,
+ 0.489,
+ 0.784,
+ 0.533
+ ],
+ "angle": 0,
+ "content": "- We formulate the image address localization problem and introduce the AddressCLIP framework for this problem by utilizing the alignment between the image and address text."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.226,
+ 0.535,
+ 0.784,
+ 0.579
+ ],
+ "angle": 0,
+ "content": "- Two key ingredients are designed for better alignment of the image and address, i.e., image-caption alignment and image-geography matching, which are mutually beneficial."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.225,
+ 0.58,
+ 0.784,
+ 0.609
+ ],
+ "angle": 0,
+ "content": "- We introduce three datasets named Pitts-IAL, SF-IAL-Base, and SF-IAL-Large to facilitate the study of the image address localization problem."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.225,
+ 0.61,
+ 0.784,
+ 0.655
+ ],
+ "angle": 0,
+ "content": "- Experiments demonstrate that our method achieves compelling performance on the proposed IAL datasets. Extensive ablations, visualizations, and analyses are provided to show the effectiveness of the proposed method."
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.225,
+ 0.489,
+ 0.784,
+ 0.655
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.688,
+ 0.388,
+ 0.705
+ ],
+ "angle": 0,
+ "content": "2 Related Work"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.72,
+ 0.788,
+ 0.842
+ ],
+ "angle": 0,
+ "content": "Image Geo-localization. Image geo-localization, or visual place recognition, is usually formulated as an image retrieval problem on the city scale, which needs to collect a geo-tagged database of pre-computed embeddings of either local or global features [6,13,24,25,33,34,40,45]. In recent years, deep learning models [14, 21, 43] have been proven to perform remarkably in image feature extraction, complemented with an aggregation or pooling layer [3,4,9,16,17,27,32,37,54]. Recent methods achieve impressive retrieval performance by performing an additional re-ranking phase [19,48,53], adopting powerful pre-trained backbones [35]"
+ }
+ ],
+ [
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.218,
+ 0.116,
+ 0.23,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "4"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.272,
+ 0.115,
+ 0.351,
+ 0.128
+ ],
+ "angle": 0,
+ "content": "S. Xu et al."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.218,
+ 0.144,
+ 0.369,
+ 0.156
+ ],
+ "angle": 0,
+ "content": "(a) Administrative Address"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.158,
+ 0.77,
+ 0.184
+ ],
+ "angle": 0,
+ "content": "Example: Carnegie Mellon University, 5000, Forbes Avenue, Oakland, Pittsburgh, Pennsylvania, USA \nHierarchy: Building Name, House Number, Street, Neighborhood, City, County/State, Country"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.218,
+ 0.194,
+ 0.39,
+ 0.206
+ ],
+ "angle": 0,
+ "content": "(b) Semantic Address Partition"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.236,
+ 0.207,
+ 0.477,
+ 0.256
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_footnote",
+ "bbox": [
+ 0.245,
+ 0.261,
+ 0.462,
+ 0.29
+ ],
+ "angle": 0,
+ "content": "located in South Side Flats, specifically on a section of East Carson Street lying between South 10th Street and South 12th Street"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.502,
+ 0.194,
+ 0.653,
+ 0.207
+ ],
+ "angle": 0,
+ "content": "(c) Vision-Language Model"
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.506,
+ 0.207,
+ 0.782,
+ 0.277
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_footnote",
+ "bbox": [
+ 0.509,
+ 0.279,
+ 0.614,
+ 0.289
+ ],
+ "angle": 0,
+ "content": "Candidate Addresses"
+ },
+ {
+ "type": "image_footnote",
+ "bbox": [
+ 0.633,
+ 0.275,
+ 0.778,
+ 0.289
+ ],
+ "angle": 0,
+ "content": "located in the Downtown, specifically on a section of Third Avenue that near Grant Street"
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.214,
+ 0.3,
+ 0.785,
+ 0.342
+ ],
+ "angle": 0,
+ "content": "Fig. 2: The problem statement of the image address localization task consists of (a) examples of administrative address and hierarchy, (b) semantic address partition strategy, and (c) address predicting using visual-language models."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.382,
+ 0.788,
+ 0.548
+ ],
+ "angle": 0,
+ "content": "to extract image features [22, 28], or training on large-scale place recognition datasets [2, 3, 7, 22]. Different from retrieval-based methods, classification-based methods focus on planet-scale localization and split the earth into disjoint regions to classify [12, 36, 42, 46, 49]. More recently, StreetCLIP [18] and GeoCLIP [10] both utilize the vision-language model CLIP [38] with region description or GPS information for better generalizability. Going beyond image geo-localization, we propose to perform image address localization to obtain readable textual addresses rather than digital coordinates without a retrieval gallery. This not only enables models to directly output human-understandable semantic addresses for a given image but also paves the way for more complex geographical human-computer interactions in the future."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.553,
+ 0.789,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "Transfer Learning in Vision-Language Models. The integration of language supervision with visual data is garnering significant interest, with the primary aim being to align images and texts and learn a shared embedding space. As outlined in [52], the advancements in vision-language models can largely be attributed to three key developments: Transformers [47], contrastive representation learning [11, 20], and expansive web-scale training datasets [26, 39]. One notable example is CLIP [38], which employs two encoder networks trained via contrastive loss to align image-text pairs, thus enabling impressive zero-shot performance. Adapting CLIP to downstream tasks typically involves either full fine-tuning or linear probing [15]. Recently, prompt learning offers an alternative by introducing a small number of trainable prompt tokens at the input. Learnable prompts can be applied to the language branch [52], image instances [51], or both forming a multi-modal prompt [29]. Complete fine-tuning enables CLIP to fully adapt to the data distribution of downstream tasks, while prompt learning enhances CLIP's zero-shot learning capabilities. Due to the domain gap between the IAL task and the pre-training tasks, our proposed AddressCLIP adopts carefully designed image-caption alignment and image-geography matching to transfer CLIP toward the address localization task, which is superior to the direct complete fine-tuning way."
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.64,
+ 0.115,
+ 0.732,
+ 0.128
+ ],
+ "angle": 0,
+ "content": "AddressCLIP"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.775,
+ 0.116,
+ 0.787,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "5"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.145,
+ 0.442,
+ 0.161
+ ],
+ "angle": 0,
+ "content": "3 Problem Statement"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.175,
+ 0.788,
+ 0.281
+ ],
+ "angle": 0,
+ "content": "In this study, we focus on the city-wide image address localization problem. The administrative address hierarchy around the world varies widely depending on the history, geography, culture, and political systems of each country. Taking the United States as an example, we provide a specific illustration of an administrative address and its corresponding hierarchy in Fig. 2 (a). Since images in one dataset belong to the same city, our study distinguishes image addresses on neighborhood and street levels."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.282,
+ 0.788,
+ 0.493
+ ],
+ "angle": 0,
+ "content": "The straightforward division mentioned above introduces two challenges in practical city-wide scenarios. Firstly, variable street lengths can result in coarsely localized addresses, particularly for highways that extend for kilometers, creating a pronounced long-tail distribution issue and diverse inner-address visual features that hinder precise localization during inference. Secondly, address ambiguity arises at street intersections, where images could be equally attributed to intersecting streets, thus lacking a clear and singular textual supervision signal. To address these concerns, we introduce a semantic address partition strategy for a more granular segmentation of streets, as shown in Fig. 2 (b). By segmenting streets at intersections, we achieve a balance in street lengths, which refines the address localization scope and eliminates the intersection ambiguity, aligning more closely with the way humans typically describe locations. In this way, the textual representation of addresses consists of the main street name (marked green) and the name of one or two streets that intersect it (marked brown)."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.493,
+ 0.789,
+ 0.6
+ ],
+ "angle": 0,
+ "content": "Formally, the Image Address Localization problem is defined as follows: given a training dataset \\( D_{train} = \\{(I_i,A_i)\\}_{i = 1}^M \\) containing pairs of image \\( I_{i} \\) and address \\( A_{i} \\), our objective is to train a vision-language model \\( \\mathcal{H}_{\\theta} \\) and use it to predict the address of query images, \\( A_{k}^{Q} = \\mathcal{H}_{\\theta}(I_{k}^{Q}),\\forall k\\in [1..K] \\) where \\( I_k^Q\\in D_{test} \\). The images in the query set \\( I^Q \\) can belong to any candidate address in the same city as the images in the training set. Fig. 2 (c) shows a schematic diagram of predicting the readable textual address for a given query image."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.62,
+ 0.383,
+ 0.636
+ ],
+ "angle": 0,
+ "content": "4 AddressCLIP"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.651,
+ 0.439,
+ 0.665
+ ],
+ "angle": 0,
+ "content": "4.1 Framework Overview"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.674,
+ 0.788,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "We formulate the IAL problem as a vision-text alignment problem between the image and address pairs. Fig. 3 depicts the framework of our method. During training, the embeddings of the image and the address are extracted by the image encoder and text encoder, respectively, and are then aligned through image-address contrastive learning. An additive scene caption is introduced as a supplement to the address to enrich the plain text information. The scene caption shares the same text encoder with the image address, and the resulting caption embedding and image embedding are combined for image-caption contrastive learning. Furthermore, we adopt the geographical position information as a guide to increase the similarities between geographically close image features while increasing the differences between geographically distant image features."
+ }
+ ],
+ [
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.218,
+ 0.116,
+ 0.23,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "6"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.272,
+ 0.115,
+ 0.351,
+ 0.128
+ ],
+ "angle": 0,
+ "content": "S. Xu et al."
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 0.218,
+ 0.144,
+ 0.789,
+ 0.352
+ ],
+ "angle": 0,
+ "content": null
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 0.214,
+ 0.36,
+ 0.788,
+ 0.417
+ ],
+ "angle": 0,
+ "content": "Fig. 3: Overview of the proposed AddressCLIP framework. (a) During training, the alignment of image and address is learned by the image-address contrastive loss, image-capture contrastive loss, and image-geography matching loss. (b) At inference, the address with the highest similarity to the query image's embedding is chosen."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.449,
+ 0.788,
+ 0.495
+ ],
+ "angle": 0,
+ "content": "The image-geography matching is learned between geospatial distance and image feature similarity. During inference, the address with the highest similarity to the query image's embedding indicates the most probable address."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.522,
+ 0.45,
+ 0.538
+ ],
+ "angle": 0,
+ "content": "4.2 Image-Text Alignment"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.552,
+ 0.788,
+ 0.688
+ ],
+ "angle": 0,
+ "content": "It is reasonable to use address information directly as textual prompts for image-address alignment learning. However, the address text is simple and limited. It cannot provide context about environments, landmarks, or other entities, which are crucial for precise address localization. To alleviate the issues, we incorporate additional descriptive captions that capture the nuances of the visual scene, thereby endowing the model with a deeper understanding of the contextual elements that are often missing in the bare address labels. This mechanism enables more accurate and context-aware predictions by effectively bridging the gap between visual perception and textual representation."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.69,
+ 0.789,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "Scene description can be generated through manual annotation, which, although accurate, is costly and not easily scalable to large datasets. Benefiting from the advancements in vision-language models, we utilize pre-trained vision-language models [30] to generate linguistic captions corresponding to image scenes. The lower left corner of Fig. 3(a) shows some illustrative examples, where the descriptions can include context like the presence of specific buildings or unique street signs, which is relevant for distinguishing between visually similar but geographically distant locations. This also aligns the model's learning process with how humans typically communicate location information. For detailed analyses of scene captions, refer to the supplementary material."
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.64,
+ 0.115,
+ 0.732,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "AddressCLIP"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.776,
+ 0.117,
+ 0.786,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "7"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.215,
+ 0.147,
+ 0.788,
+ 0.268
+ ],
+ "angle": 0,
+ "content": "Formally, define image features extracted from the image encoder \\(\\mathcal{V}(\\cdot)\\) as \\(V_{i} = \\mathcal{V}(I_{i}),\\forall i\\in [1,\\dots,N]\\). The text encoder \\(\\mathcal{V}(\\cdot)\\) extracts address features \\(T_{i}^{A} = \\mathcal{T}(A_{i})\\) and caption features \\(T_{i}^{C} = \\mathcal{T}(C_{i} + A_{i})\\), where the scene caption \\(C_i\\) is obtained by a vision-language model. We experimentally observe that appending address information to the scene caption is more conducive to address localization, which is discussed in detail in Sec. 6.3. Note that the additive scene caption is only used for training. The alignment of images and addresses is learned via image-address contrastive loss and image-caption contrastive loss."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.215,
+ 0.268,
+ 0.788,
+ 0.299
+ ],
+ "angle": 0,
+ "content": "For a batch of size \\( N \\) comprising image-text pairs, the image-address contrastive loss can be written as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.239,
+ 0.304,
+ 0.786,
+ 0.342
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} _ {\\text {a d d r e s s}} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left[ \\log \\frac {\\exp \\left(V _ {i} \\cdot T _ {i} ^ {A} / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(V _ {i} \\cdot T _ {j} ^ {A} / \\tau\\right)} + \\log \\frac {\\exp \\left(T _ {i} ^ {A} \\cdot V _ {i} / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(T _ {i} ^ {A} \\cdot V _ {k} / \\tau\\right)} \\right], \\tag {1}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.215,
+ 0.348,
+ 0.788,
+ 0.378
+ ],
+ "angle": 0,
+ "content": "where \\(\\tau\\) is the temperature parameter. Similarly, the image-caption contrastive loss is formulated as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.239,
+ 0.383,
+ 0.786,
+ 0.422
+ ],
+ "angle": 0,
+ "content": "\\[\n\\mathcal {L} _ {c a p t i o n} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left[ \\log \\frac {\\exp (V _ {i} \\cdot T _ {i} ^ {C} / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (V _ {i} \\cdot T _ {j} ^ {C} / \\tau)} + \\log \\frac {\\exp (T _ {i} ^ {C} \\cdot V _ {i} / \\tau)}{\\sum_ {k = 1} ^ {N} \\exp (T _ {i} ^ {C} \\cdot V _ {k} / \\tau)} \\right]. \\quad (2)\n\\]"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.439,
+ 0.494,
+ 0.455
+ ],
+ "angle": 0,
+ "content": "4.3 Image-Geography Matching"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.215,
+ 0.462,
+ 0.788,
+ 0.628
+ ],
+ "angle": 0,
+ "content": "In general, address text in city-wide scenarios may be geographically far away but highly similar, or geographically close but significantly different. This makes image-address alignment learning difficult to optimize with address text alone. In contrast, the geographic coordinates of images (e.g., UTM coordinates) differ significantly, showcasing clear distinctions and discriminative properties. From the perspective of manifold learning, image embedding represents a low-dimensional representation of images in the feature space, and its distribution should be consistent with the geographic coordinates of the images. Our goal is to ensure that geographically proximate images exhibit closely in the feature space, while geographically distant images reflect more within the feature space. Visualization results and analysis are elaborated in Sec. 6.4."
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.215,
+ 0.629,
+ 0.788,
+ 0.72
+ ],
+ "angle": 0,
+ "content": "Inspired by the above motivation, we propose an image-geography matching loss to constrain image features according to the spatial distances of geographic coordinates. Specifically, denote \\( U_{i}:\\mathbb{UTM}_{east}\\times \\mathbb{UTM}_{north},\\forall i\\in [1,\\dots ,N] \\) the set of geographic coordinates corresponding to all images within a batch of size \\( N \\). We can calculate each element of the spatial distance matrix \\( D^{U} \\) in the geographic space as follows:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.327,
+ 0.728,
+ 0.786,
+ 0.761
+ ],
+ "angle": 0,
+ "content": "\\[\nD _ {i j} ^ {U} = \\left\\| \\hat {U} _ {i} - \\hat {U} _ {j} \\right\\| _ {1}, s. t., \\hat {U} _ {i} = \\frac {U _ {i} - \\operatorname* {m i n} \\left(U _ {i}\\right)}{\\operatorname* {m a x} \\left(U _ {i}\\right) - \\operatorname* {m i n} \\left(U _ {i}\\right)}, \\tag {3}\n\\]"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.215,
+ 0.769,
+ 0.788,
+ 0.814
+ ],
+ "angle": 0,
+ "content": "where Manhattan distance and min-max normalization are adopted. Correspondingly, each element of the feature similarity matrix \\( D^{V} \\) in the image embedding space is calculated as:"
+ },
+ {
+ "type": "equation",
+ "bbox": [
+ 0.435,
+ 0.812,
+ 0.786,
+ 0.845
+ ],
+ "angle": 0,
+ "content": "\\[\nD _ {i j} ^ {V} = \\frac {V _ {i} \\cdot V _ {j}}{| | V _ {i} | | \\cdot | | V _ {j} | |}. \\tag {4}\n\\]"
+ }
+ ],
+ [
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.218,
+ 0.116,
+ 0.23,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "8"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.272,
+ 0.115,
+ 0.351,
+ 0.128
+ ],
+ "angle": 0,
+ "content": "S. Xu et al."
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 0.219,
+ 0.148,
+ 0.782,
+ 0.162
+ ],
+ "angle": 0,
+ "content": "Table 1: Detailed information of the proposed Image Address Localization datasets."
+ },
+ {
+ "type": "table",
+ "bbox": [
+ 0.236,
+ 0.163,
+ 0.769,
+ 0.257
+ ],
+ "angle": 0,
+ "content": "| Visual input examples for Image Address Localization: |
| Grant Street, Downtown | Fifth Avenue, Downtown |
| User | Where might this photo have been taken? Tell me its street level address. |
| LLaVA-IAL | The address of this photo might be Grant Street, Downtown, Pittsburgh, PA, USA. | The address of this photo might be Fifth Avenue, Downtown, Pittsburgh, PA, USA. |
| GPT-4V [1] | This photo was taken in Peavey Plaza in Minneapolis, Minnesota, USA, on the corner of Nicolson Mall and 11th Street. | This photo was taken at 400 Fifth Avenue in Pittsburgh, Pennsylvania. |
| Gemini-Pro-V [44] | This photo was taken on the Rose Fitzgerald Kennedy Greenway in Boston, Massachusetts. | The photo was taken on Fifth Avenue in Pittsburgh, Pennsylvania. |
| QWen-VL [5] | This photo appears to have been taken on a city street, possibly in a downtown area or commercial district. | This photo appears to have been taken on Fifth Avenue in Pittsburgh, Pennsylvania, USA. The street sign in the image confirms this location. |
"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.423,
+ 0.505,
+ 0.437
+ ],
+ "angle": 0,
+ "content": "6.6 Limitations and Future Work"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.45,
+ 0.788,
+ 0.631
+ ],
+ "angle": 0,
+ "content": "The proposed AddressCLIP can be regarded as a discriminative model, limited by the set of candidate addresses at inference. In contrast, generative models such as multimodal large language models (MLLMs) [1, 5, 44] can yield more flexible and interactive geographic textual information but may not offer precise administrative addresses. In future work, we plan to explore the potential of MLLMs for the IAL task. To this end, we have made an attempt by constructing a question-and-answer dataset based on Pitts-IAL and adopting the LLaVA-1.5 [31] model for instruct tuning. Tab. 7 shows some examples comparing with representative MLLMs. It is evident that LLaVA-IAL, fine-tuned with instructions, can predict accurate administrative addresses consistent with address hierarchy, while other compared MLLMs are unable to predict addresses without landmarks or street signs and lack standardized output formats."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.216,
+ 0.656,
+ 0.36,
+ 0.672
+ ],
+ "angle": 0,
+ "content": "7 Conclusion"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.214,
+ 0.689,
+ 0.788,
+ 0.841
+ ],
+ "angle": 0,
+ "content": "In this study, we introduce the problem of image address localization and propose three IAL datasets for evaluation and subsequent research. To facilitate the alignment of images and addresses for tackling the problem, we propose the AddressCLIP framework consisting of image-text alignment and image-geography matching. Extensive experiments on the proposed datasets validate that our method outperforms transfer learning methods that transfer CLIP to downstream tasks. We compare the proposed method with the existing two-stage address localization pipeline based on the image geo-localization technology and discuss AddressCLIP's application in real-world situations. Finally, we explore the potential of multimodal large language models for address localization."
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.64,
+ 0.115,
+ 0.732,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "AddressCLIP"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.769,
+ 0.117,
+ 0.786,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "15"
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.218,
+ 0.145,
+ 0.403,
+ 0.163
+ ],
+ "angle": 0,
+ "content": "Acknowledgements"
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 0.217,
+ 0.179,
+ 0.785,
+ 0.21
+ ],
+ "angle": 0,
+ "content": "This work was supported by the National Natural Science Foundations of China (Grants No.62376267, 62076242) and the innoHK project."
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 0.218,
+ 0.234,
+ 0.323,
+ 0.25
+ ],
+ "angle": 0,
+ "content": "References"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.267,
+ 0.786,
+ 0.309
+ ],
+ "angle": 0,
+ "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.311,
+ 0.786,
+ 0.337
+ ],
+ "angle": 0,
+ "content": "2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.339,
+ 0.786,
+ 0.365
+ ],
+ "angle": 0,
+ "content": "3. Ali-Bey, A., Chaib-Draa, B., Giguere, P.: Mixvpr: Feature mixing for visual place recognition. In: WACV. pp. 2998-3007 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.367,
+ 0.786,
+ 0.393
+ ],
+ "angle": 0,
+ "content": "4. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: CVPR. pp. 5297-5307 (2016)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.395,
+ 0.786,
+ 0.435
+ ],
+ "angle": 0,
+ "content": "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.437,
+ 0.786,
+ 0.463
+ ],
+ "angle": 0,
+ "content": "6. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer vision and image understanding 110(3), 346-359 (2008)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.465,
+ 0.786,
+ 0.491
+ ],
+ "angle": 0,
+ "content": "7. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: CVPR. pp. 4878-4888 (2022)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.493,
+ 0.786,
+ 0.546
+ ],
+ "angle": 0,
+ "content": "8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.548,
+ 0.786,
+ 0.575
+ ],
+ "angle": 0,
+ "content": "9. Berton, G.M., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: WACV. pp. 2918-2927 (2021)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.577,
+ 0.786,
+ 0.616
+ ],
+ "angle": 0,
+ "content": "0. Cepeda, V.V., Nayak, G.K., Shah, M.: Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. arXiv preprint arXiv:2309.16020 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.618,
+ 0.786,
+ 0.645
+ ],
+ "angle": 0,
+ "content": "1. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.647,
+ 0.786,
+ 0.686
+ ],
+ "angle": 0,
+ "content": "2. Clark, B., Kerrigan, A., Kulkarni, P.P., Cepeda, V.V., Shah, M.: Where we are and what we're looking at: Query based worldwide image geo-localization using hierarchies and scenes. In: CVPR. pp. 23182-23190 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.688,
+ 0.786,
+ 0.714
+ ],
+ "angle": 0,
+ "content": "3. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: ECCV Workshop. vol. 1, pp. 1-2 (2004)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.716,
+ 0.786,
+ 0.77
+ ],
+ "angle": 0,
+ "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.772,
+ 0.786,
+ 0.812
+ ],
+ "angle": 0,
+ "content": "5. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.226,
+ 0.814,
+ 0.786,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "6. Ge, Y., Wang, H., Zhu, F., Zhao, R., Li, H.: Self-supervising fine-grained region similarities for large-scale image localization. In: ECCV. pp. 369-386 (2020)"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.226,
+ 0.267,
+ 0.786,
+ 0.84
+ ],
+ "angle": 0,
+ "content": null
+ }
+ ],
+ [
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.218,
+ 0.116,
+ 0.236,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "16"
+ },
+ {
+ "type": "header",
+ "bbox": [
+ 0.272,
+ 0.115,
+ 0.351,
+ 0.128
+ ],
+ "angle": 0,
+ "content": "S. Xu et al."
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.218,
+ 0.148,
+ 0.785,
+ 0.189
+ ],
+ "angle": 0,
+ "content": "17. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: End-to-end learning of deep visual representations for image retrieval. International Journal of Computer Vision 124(2), 237-254 (2017)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.218,
+ 0.19,
+ 0.785,
+ 0.217
+ ],
+ "angle": 0,
+ "content": "18. Haas, L., Alberti, S., Skreta, M.: Learning generalized zero-shot learners for open-domain image geolocation. arXiv preprint arXiv:2302.00275 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.218,
+ 0.218,
+ 0.785,
+ 0.257
+ ],
+ "angle": 0,
+ "content": "19. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: CVPR. pp. 14141-14152 (2021)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.258,
+ 0.785,
+ 0.284
+ ],
+ "angle": 0,
+ "content": "20. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.285,
+ 0.785,
+ 0.311
+ ],
+ "angle": 0,
+ "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.312,
+ 0.785,
+ 0.339
+ ],
+ "angle": 0,
+ "content": "22. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.34,
+ 0.785,
+ 0.366
+ ],
+ "angle": 0,
+ "content": "23. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.367,
+ 0.785,
+ 0.392
+ ],
+ "angle": 0,
+ "content": "24. Jegou, H., Douze, M., Schmid, C.: Hamming embedding and weak geometric consistency for large scale image search. In: ECCV. pp. 304-317 (2008)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.393,
+ 0.785,
+ 0.433
+ ],
+ "angle": 0,
+ "content": "25. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE transactions on pattern analysis and machine intelligence 34(9), 1704-1716 (2011)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.434,
+ 0.785,
+ 0.473
+ ],
+ "angle": 0,
+ "content": "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.475,
+ 0.785,
+ 0.5
+ ],
+ "angle": 0,
+ "content": "27. Jin Kim, H., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geo-localization. In: CVPR. pp. 2136-2145 (2017)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.501,
+ 0.785,
+ 0.542
+ ],
+ "angle": 0,
+ "content": "28. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.543,
+ 0.785,
+ 0.569
+ ],
+ "angle": 0,
+ "content": "29. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR. pp. 19113-19122 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.57,
+ 0.785,
+ 0.609
+ ],
+ "angle": 0,
+ "content": "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900 (2022)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.61,
+ 0.785,
+ 0.637
+ ],
+ "angle": 0,
+ "content": "31. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.638,
+ 0.785,
+ 0.664
+ ],
+ "angle": 0,
+ "content": "32. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: ICCV. pp. 2570-2579 (2019)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.665,
+ 0.785,
+ 0.691
+ ],
+ "angle": 0,
+ "content": "33. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.692,
+ 0.785,
+ 0.718
+ ],
+ "angle": 0,
+ "content": "34. Oliva, A., Torralba, A.: Building the gist of a scene: The role of global image features in recognition. Progress in brain research 155, 23-36 (2006)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.719,
+ 0.785,
+ 0.758
+ ],
+ "angle": 0,
+ "content": "35. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.759,
+ 0.785,
+ 0.8
+ ],
+ "angle": 0,
+ "content": "36. Pramanick, S., Nowara, E.M., Gleason, J., Castillo, C.D., Chellappa, R.: Where in the world is this image? transformer-based geo-localization in the wild. In: ECCV. pp. 196-215 (2022)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.217,
+ 0.801,
+ 0.785,
+ 0.84
+ ],
+ "angle": 0,
+ "content": "37. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE transactions on pattern analysis and machine intelligence 41(7), 1655-1668 (2018)"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.217,
+ 0.148,
+ 0.785,
+ 0.84
+ ],
+ "angle": 0,
+ "content": null
+ }
+ ],
+ [
+ {
+ "type": "header",
+ "bbox": [
+ 0.64,
+ 0.115,
+ 0.732,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "AddressCLIP"
+ },
+ {
+ "type": "page_number",
+ "bbox": [
+ 0.769,
+ 0.116,
+ 0.786,
+ 0.127
+ ],
+ "angle": 0,
+ "content": "17"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.214,
+ 0.147,
+ 0.788,
+ 0.189
+ ],
+ "angle": 0,
+ "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.214,
+ 0.19,
+ 0.788,
+ 0.231
+ ],
+ "angle": 0,
+ "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.214,
+ 0.232,
+ 0.787,
+ 0.259
+ ],
+ "angle": 0,
+ "content": "40. Schindler, G., Brown, M., Szeliski, R.: City-scale location recognition. In: CVPR. pp. 1-7. IEEE (2007)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.26,
+ 0.787,
+ 0.34
+ ],
+ "angle": 0,
+ "content": "41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv abs/2210.08402 (2022), https://api_semanticscholar.org/CorpusID: 252917726"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.342,
+ 0.787,
+ 0.37
+ ],
+ "angle": 0,
+ "content": "42. Seo, P.H., Weyand, T., Sim, J., Han, B.: Cplanet: Enhancing image geolocation by combinatorial partitioning of maps. In: ECCV. pp. 536-551 (2018)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.371,
+ 0.787,
+ 0.397
+ ],
+ "angle": 0,
+ "content": "43. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.398,
+ 0.787,
+ 0.438
+ ],
+ "angle": 0,
+ "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.439,
+ 0.787,
+ 0.466
+ ],
+ "angle": 0,
+ "content": "45. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: CVPR. pp. 883-890 (2013)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.467,
+ 0.787,
+ 0.508
+ ],
+ "angle": 0,
+ "content": "46. Trivigno, G., Berton, G., Aragon, J., Caputo, B., Masone, C.: Divide&classify: Fine-grained classification for city-wide visual geo-localization. In: ICCV. pp. 11142-11152 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.508,
+ 0.787,
+ 0.536
+ ],
+ "angle": 0,
+ "content": "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.537,
+ 0.787,
+ 0.577
+ ],
+ "angle": 0,
+ "content": "48. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: CVPR. pp. 13648-13657 (2022)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.578,
+ 0.787,
+ 0.605
+ ],
+ "angle": 0,
+ "content": "49. Weyand, T., Kostrikov, I., Philbin, J.: Planet-photo geolocation with convolutional neural networks. In: ECCV. pp. 37-55 (2016)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.606,
+ 0.787,
+ 0.633
+ ],
+ "angle": 0,
+ "content": "50. Wilson, D., Zhang, X., Sultani, W., Wshah, S.: Image and object geo-localization. International Journal of Computer Vision pp. 1-43 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.634,
+ 0.787,
+ 0.66
+ ],
+ "angle": 0,
+ "content": "51. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.661,
+ 0.787,
+ 0.688
+ ],
+ "angle": 0,
+ "content": "52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.689,
+ 0.787,
+ 0.729
+ ],
+ "angle": 0,
+ "content": "53. Zhu, S., Yang, L., Chen, C., Shah, M., Shen, X., Wang, H.: R2former: Unified retrieval and reranking transformer for place recognition. In: CVPR. pp. 19370-19380 (2023)"
+ },
+ {
+ "type": "ref_text",
+ "bbox": [
+ 0.215,
+ 0.73,
+ 0.787,
+ 0.757
+ ],
+ "angle": 0,
+ "content": "54. Zhu, Y., Wang, J., Xie, L., Zheng, L.: Attention-based pyramid aggregation network for visual place recognition. In: ACM MM. pp. 99-107 (2018)"
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 0.214,
+ 0.147,
+ 0.788,
+ 0.757
+ ],
+ "angle": 0,
+ "content": null
+ }
+ ]
+]
\ No newline at end of file
diff --git a/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/bbd4751a-6e05-4f3a-882f-4ee98bbd88b8_origin.pdf b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/bbd4751a-6e05-4f3a-882f-4ee98bbd88b8_origin.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..e450371b6941763c569788dbfab7e99ddccddf20
--- /dev/null
+++ b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/bbd4751a-6e05-4f3a-882f-4ee98bbd88b8_origin.pdf
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2548c59a991ebe01d5b13fd77c3eb5e7dd5ba7e616665d9413a399dd25b5109
+size 4446257
diff --git a/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/full.md b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/full.md
new file mode 100644
index 0000000000000000000000000000000000000000..831fbf9dbb9fc922d9f63d6a45fcdbfc98954a3c
--- /dev/null
+++ b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/full.md
@@ -0,0 +1,331 @@
+# AddressCLIP: Empowering Vision-Language Models for City-wide Image Address Localization
+
+Shixiong $\mathrm{Xu}^{1,3*†}$ , Chenghao Zhang $^{2*}$ , Lubin Fan $^{2‡}$ , Gaofeng Meng $^{1,3,4‡}$ , Shiming Xiang $^{1,3}$ , and Jieping Ye $^{2}$
+
+$^{1}$ MAIS, Institute of Automation, Chinese Academy of Sciences $^{2}$ Alibaba Cloud
+
+$^{3}$ School of Artificial Intelligence, University of Chinese Academy of Sciences
+$^{4}$ CAIR, HK Institute of Science & Innovation, Chinese Academy of Sciences
+
+Abstract. In this study, we introduce a new problem raised by social media and photojournalism, named Image Address Localization (IAL), which aims to predict the readable textual address where an image was taken. Existing two-stage approaches involve predicting geographical coordinates and converting them into human-readable addresses, which can lead to ambiguity and be resource-intensive. In contrast, we propose an end-to-end framework named AddressCLIP to solve the problem with more semantics, consisting of two key ingredients: i) image-text alignment to align images with addresses and scene captions by contrastive learning, and ii) image-geography matching to constrain image features with the spatial distance in terms of manifold learning. Additionally, we have built three datasets from Pittsburgh and San Francisco on different scales specifically for the IAL problem. Experiments demonstrate that our approach achieves compelling performance on the proposed datasets and outperforms representative transfer learning methods for vision-language models. Furthermore, extensive ablations and visualizations exhibit the effectiveness of the proposed method. The datasets and source code are available at https://github.com/xsx1001/AddressCLIP.
+
+Keywords: Image address localization $\cdot$ Image-text alignment $\cdot$ Image-geography matching $\cdot$ Vision-language model
+
+# 1 Introduction
+
+Users on social media platforms such as Facebook and Instagram often tag their pictures with textual addresses to connect with local communities, raising the demand for predicting the descriptive address information of the place where an image was taken. This has various practical applications, for instance, businesses and travel platforms can use addresses of images to provide recommendations
+
+
+Fig. 1: Comparison of image-based geo-localization and address localization tasks. The objective of the proposed task is to predict the semantic text address of a given image instead of a digital GPS coordinate without the need for a retrieval gallery.
+
+
+
+or organize location-specific content. Additionally, photojournalism can rapidly verify the authenticity of the event with the image's address.
+
+To predict an image's address, one reasonable approach involves leveraging image geo-localization technology to predict GPS coordinates (i.e., latitude and longitude) from an image [50], followed by the reverse Geocoding to query for a readable address. Image geo-localization, also known as visual place recognition, is commonly treated as an image retrieval problem where a database of geo-tagged images serves as a matching reference for the query image. Previous retrieval-based methods [3,4,7,19,28] have shown remarkable performance. However, in practice, the creation of pre-collected geo-tagged databases requires significant labor and storage resources, while GPS coordinates lack readability and semantics. In addition, the conversion from GPS to readable addresses often presents ambiguities, and the Image-GPS-Address pipeline is not end-to-end.
+
+To alleviate the above issues, in this study, we propose to perform Image Address Localization (IAL) where a model is tasked to predict the readable textual address where a given image was taken. We design a semantic address partition strategy to perform fine-grained partitioning of city-wide addresses, conforming to the way humans describe address information. By doing this, we are able to train models in an end-to-end manner, and during inference, there is no need to construct a retrieval database which greatly reduces the storage and retrieval burden. Furthermore, the model's output addresses align more closely with human description habits, which provides a bridge for subsequent city-wide scene understanding and point-of-interest recommendation. Fig. 1 shows the comparison of image geo-localization and image address localization tasks, where the latter focuses on predicting human-readable textual address information.
+
+In this study, we propose an end-to-end framework, AddressCLIP, based on the visual-language model CLIP [38], aiming to learn an alignment of images and addresses. Our approach leverages two key ingredients: image-text alignment and image-geography matching. Firstly, we introduce additional scene captions as a supplement to address text thus facilitating the alignment of images and tex
+
+tual addresses by contrastive learning. Secondly, we propose an image-geography matching mechanism to bring features of geographically proximate images closer while separating features of images that are far apart geographically.
+
+To support the image address localization task, we constructed three IAL datasets of different sizes based on the Pitts-250k [4] and SF-XL datasets [7]: Pitts-IAL (234K), SF-IAL-Base (184K), and SF-IAL-Large (1.96M). In contrast to the original datasets, each image in our dataset is accompanied by not only its geographical coordinate but also the administrative address. Specifically, we utilized the reverse Geocoding API of Google Maps to retrieve administrative addresses for a portion of the images and obtain addresses for the remaining images through nearest-neighbor interpolation of the geographical coordinates.
+
+We evaluate the proposed AddressCLIP framework on the introduced datasets. Our proposed method achieves a Top-1 address localization accuracy of over $80\%$ across three IAL datasets, most notably reaching a performance of $85.92\%$ on the largest dataset, SF-IAL-Large. Compared with challenging baselines [29, 51, 52] that transfer CLIP to the downstream IAL task, our AddressCLIP achieves improvements of $3\%$ to $6\%$ on the proposed datasets. In addition, the qualitative results demonstrate good alignment between images and textual address queries in geographical space. Finally, we discuss the superiority of the proposed method over the two-stage "Image-GPS-Address" approach and explore the application prospects of multimodal large language models in the IAL task.
+
+Our contributions are summarized as follows:
+
+- We formulate the image address localization problem and introduce the AddressCLIP framework for this problem by utilizing the alignment between the image and address text.
+- Two key ingredients are designed for better alignment of the image and address, i.e., image-caption alignment and image-geography matching, which are mutually beneficial.
+- We introduce three datasets named Pitts-IAL, SF-IAL-Base, and SF-IAL-Large to facilitate the study of the image address localization problem.
+- Experiments demonstrate that our method achieves compelling performance on the proposed IAL datasets. Extensive ablations, visualizations, and analyses are provided to show the effectiveness of the proposed method.
+
+# 2 Related Work
+
+Image Geo-localization. Image geo-localization, or visual place recognition, is usually formulated as an image retrieval problem on the city scale, which needs to collect a geo-tagged database of pre-computed embeddings of either local or global features [6,13,24,25,33,34,40,45]. In recent years, deep learning models [14, 21, 43] have been proven to perform remarkably in image feature extraction, complemented with an aggregation or pooling layer [3,4,9,16,17,27,32,37,54]. Recent methods achieve impressive retrieval performance by performing an additional re-ranking phase [19,48,53], adopting powerful pre-trained backbones [35]
+
+(a) Administrative Address
+
+Example: Carnegie Mellon University, 5000, Forbes Avenue, Oakland, Pittsburgh, Pennsylvania, USA
+Hierarchy: Building Name, House Number, Street, Neighborhood, City, County/State, Country
+
+(b) Semantic Address Partition
+
+located in South Side Flats, specifically on a section of East Carson Street lying between South 10th Street and South 12th Street
+
+(c) Vision-Language Model
+Fig. 2: The problem statement of the image address localization task consists of (a) examples of administrative address and hierarchy, (b) semantic address partition strategy, and (c) address predicting using visual-language models.
+
+Candidate Addresses
+located in the Downtown, specifically on a section of Third Avenue that near Grant Street
+
+to extract image features [22, 28], or training on large-scale place recognition datasets [2, 3, 7, 22]. Different from retrieval-based methods, classification-based methods focus on planet-scale localization and split the earth into disjoint regions to classify [12, 36, 42, 46, 49]. More recently, StreetCLIP [18] and GeoCLIP [10] both utilize the vision-language model CLIP [38] with region description or GPS information for better generalizability. Going beyond image geo-localization, we propose to perform image address localization to obtain readable textual addresses rather than digital coordinates without a retrieval gallery. This not only enables models to directly output human-understandable semantic addresses for a given image but also paves the way for more complex geographical human-computer interactions in the future.
+
+Transfer Learning in Vision-Language Models. The integration of language supervision with visual data is garnering significant interest, with the primary aim being to align images and texts and learn a shared embedding space. As outlined in [52], the advancements in vision-language models can largely be attributed to three key developments: Transformers [47], contrastive representation learning [11, 20], and expansive web-scale training datasets [26, 39]. One notable example is CLIP [38], which employs two encoder networks trained via contrastive loss to align image-text pairs, thus enabling impressive zero-shot performance. Adapting CLIP to downstream tasks typically involves either full fine-tuning or linear probing [15]. Recently, prompt learning offers an alternative by introducing a small number of trainable prompt tokens at the input. Learnable prompts can be applied to the language branch [52], image instances [51], or both forming a multi-modal prompt [29]. Complete fine-tuning enables CLIP to fully adapt to the data distribution of downstream tasks, while prompt learning enhances CLIP's zero-shot learning capabilities. Due to the domain gap between the IAL task and the pre-training tasks, our proposed AddressCLIP adopts carefully designed image-caption alignment and image-geography matching to transfer CLIP toward the address localization task, which is superior to the direct complete fine-tuning way.
+
+# 3 Problem Statement
+
+In this study, we focus on the city-wide image address localization problem. The administrative address hierarchy around the world varies widely depending on the history, geography, culture, and political systems of each country. Taking the United States as an example, we provide a specific illustration of an administrative address and its corresponding hierarchy in Fig. 2 (a). Since images in one dataset belong to the same city, our study distinguishes image addresses on neighborhood and street levels.
+
+The straightforward division mentioned above introduces two challenges in practical city-wide scenarios. Firstly, variable street lengths can result in coarsely localized addresses, particularly for highways that extend for kilometers, creating a pronounced long-tail distribution issue and diverse inner-address visual features that hinder precise localization during inference. Secondly, address ambiguity arises at street intersections, where images could be equally attributed to intersecting streets, thus lacking a clear and singular textual supervision signal. To address these concerns, we introduce a semantic address partition strategy for a more granular segmentation of streets, as shown in Fig. 2 (b). By segmenting streets at intersections, we achieve a balance in street lengths, which refines the address localization scope and eliminates the intersection ambiguity, aligning more closely with the way humans typically describe locations. In this way, the textual representation of addresses consists of the main street name (marked green) and the name of one or two streets that intersect it (marked brown).
+
+Formally, the Image Address Localization problem is defined as follows: given a training dataset $D_{train} = \{(I_i,A_i)\}_{i = 1}^M$ containing pairs of image $I_{i}$ and address $A_{i}$ , our objective is to train a vision-language model $\mathcal{H}_{\theta}$ and use it to predict the address of query images, $A_{k}^{Q} = \mathcal{H}_{\theta}(I_{k}^{Q}),\forall k\in [1..K]$ where $I_k^Q\in D_{test}$ . The images in the query set $I^Q$ can belong to any candidate address in the same city as the images in the training set. Fig. 2 (c) shows a schematic diagram of predicting the readable textual address for a given query image.
+
+# 4 AddressCLIP
+
+# 4.1 Framework Overview
+
+We formulate the IAL problem as a vision-text alignment problem between the image and address pairs. Fig. 3 depicts the framework of our method. During training, the embeddings of the image and the address are extracted by the image encoder and text encoder, respectively, and are then aligned through image-address contrastive learning. An additive scene caption is introduced as a supplement to the address to enrich the plain text information. The scene caption shares the same text encoder with the image address, and the resulting caption embedding and image embedding are combined for image-caption contrastive learning. Furthermore, we adopt the geographical position information as a guide to increase the similarities between geographically close image features while increasing the differences between geographically distant image features.
+
+
+Fig. 3: Overview of the proposed AddressCLIP framework. (a) During training, the alignment of image and address is learned by the image-address contrastive loss, image-capture contrastive loss, and image-geography matching loss. (b) At inference, the address with the highest similarity to the query image's embedding is chosen.
+
+The image-geography matching is learned between geospatial distance and image feature similarity. During inference, the address with the highest similarity to the query image's embedding indicates the most probable address.
+
+# 4.2 Image-Text Alignment
+
+It is reasonable to use address information directly as textual prompts for image-address alignment learning. However, the address text is simple and limited. It cannot provide context about environments, landmarks, or other entities, which are crucial for precise address localization. To alleviate the issues, we incorporate additional descriptive captions that capture the nuances of the visual scene, thereby endowing the model with a deeper understanding of the contextual elements that are often missing in the bare address labels. This mechanism enables more accurate and context-aware predictions by effectively bridging the gap between visual perception and textual representation.
+
+Scene description can be generated through manual annotation, which, although accurate, is costly and not easily scalable to large datasets. Benefiting from the advancements in vision-language models, we utilize pre-trained vision-language models [30] to generate linguistic captions corresponding to image scenes. The lower left corner of Fig. 3(a) shows some illustrative examples, where the descriptions can include context like the presence of specific buildings or unique street signs, which is relevant for distinguishing between visually similar but geographically distant locations. This also aligns the model's learning process with how humans typically communicate location information. For detailed analyses of scene captions, refer to the supplementary material.
+
+Formally, define image features extracted from the image encoder $\mathcal{V}(\cdot)$ as $V_{i} = \mathcal{V}(I_{i}),\forall i\in [1,\dots,N]$ . The text encoder $\mathcal{V}(\cdot)$ extracts address features $T_{i}^{A} = \mathcal{T}(A_{i})$ and caption features $T_{i}^{C} = \mathcal{T}(C_{i} + A_{i})$ , where the scene caption $C_i$ is obtained by a vision-language model. We experimentally observe that appending address information to the scene caption is more conducive to address localization, which is discussed in detail in Sec. 6.3. Note that the additive scene caption is only used for training. The alignment of images and addresses is learned via image-address contrastive loss and image-caption contrastive loss.
+
+For a batch of size $N$ comprising image-text pairs, the image-address contrastive loss can be written as:
+
+$$
+\mathcal {L} _ {\text {a d d r e s s}} = - \frac {1}{2 N} \sum_ {i = 1} ^ {N} \left[ \log \frac {\exp \left(V _ {i} \cdot T _ {i} ^ {A} / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(V _ {i} \cdot T _ {j} ^ {A} / \tau\right)} + \log \frac {\exp \left(T _ {i} ^ {A} \cdot V _ {i} / \tau\right)}{\sum_ {k = 1} ^ {N} \exp \left(T _ {i} ^ {A} \cdot V _ {k} / \tau\right)} \right], \tag {1}
+$$
+
+where $\tau$ is the temperature parameter. Similarly, the image-caption contrastive loss is formulated as:
+
+$$
+\mathcal {L} _ {c a p t i o n} = - \frac {1}{2 N} \sum_ {i = 1} ^ {N} \left[ \log \frac {\exp (V _ {i} \cdot T _ {i} ^ {C} / \tau)}{\sum_ {j = 1} ^ {N} \exp (V _ {i} \cdot T _ {j} ^ {C} / \tau)} + \log \frac {\exp (T _ {i} ^ {C} \cdot V _ {i} / \tau)}{\sum_ {k = 1} ^ {N} \exp (T _ {i} ^ {C} \cdot V _ {k} / \tau)} \right]. \quad (2)
+$$
+
+# 4.3 Image-Geography Matching
+
+In general, address text in city-wide scenarios may be geographically far away but highly similar, or geographically close but significantly different. This makes image-address alignment learning difficult to optimize with address text alone. In contrast, the geographic coordinates of images (e.g., UTM coordinates) differ significantly, showcasing clear distinctions and discriminative properties. From the perspective of manifold learning, image embedding represents a low-dimensional representation of images in the feature space, and its distribution should be consistent with the geographic coordinates of the images. Our goal is to ensure that geographically proximate images exhibit closely in the feature space, while geographically distant images reflect more within the feature space. Visualization results and analysis are elaborated in Sec. 6.4.
+
+Inspired by the above motivation, we propose an image-geography matching loss to constrain image features according to the spatial distances of geographic coordinates. Specifically, denote $U_{i}:\mathbb{UTM}_{east}\times \mathbb{UTM}_{north},\forall i\in [1,\dots ,N]$ the set of geographic coordinates corresponding to all images within a batch of size $N$ . We can calculate each element of the spatial distance matrix $D^{U}$ in the geographic space as follows:
+
+$$
+D _ {i j} ^ {U} = \left\| \hat {U} _ {i} - \hat {U} _ {j} \right\| _ {1}, s. t., \hat {U} _ {i} = \frac {U _ {i} - \operatorname* {m i n} \left(U _ {i}\right)}{\operatorname* {m a x} \left(U _ {i}\right) - \operatorname* {m i n} \left(U _ {i}\right)}, \tag {3}
+$$
+
+where Manhattan distance and min-max normalization are adopted. Correspondingly, each element of the feature similarity matrix $D^{V}$ in the image embedding space is calculated as:
+
+$$
+D _ {i j} ^ {V} = \frac {V _ {i} \cdot V _ {j}}{| | V _ {i} | | \cdot | | V _ {j} | |}. \tag {4}
+$$
+
+Table 1: Detailed information of the proposed Image Address Localization datasets.
+
+| Visual input examples for Image Address Localization: |
| Grant Street, Downtown | Fifth Avenue, Downtown |
| User | Where might this photo have been taken? Tell me its street level address. |
| LLaVA-IAL | The address of this photo might be Grant Street, Downtown, Pittsburgh, PA, USA. | The address of this photo might be Fifth Avenue, Downtown, Pittsburgh, PA, USA. |
| GPT-4V [1] | This photo was taken in Peavey Plaza in Minneapolis, Minnesota, USA, on the corner of Nicolson Mall and 11th Street. | This photo was taken at 400 Fifth Avenue in Pittsburgh, Pennsylvania. |
| Gemini-Pro-V [44] | This photo was taken on the Rose Fitzgerald Kennedy Greenway in Boston, Massachusetts. | The photo was taken on Fifth Avenue in Pittsburgh, Pennsylvania. |
| QWen-VL [5] | This photo appears to have been taken on a city street, possibly in a downtown area or commercial district. | This photo appears to have been taken on Fifth Avenue in Pittsburgh, Pennsylvania, USA. The street sign in the image confirms this location. |
+
+# 6.6 Limitations and Future Work
+
+The proposed AddressCLIP can be regarded as a discriminative model, limited by the set of candidate addresses at inference. In contrast, generative models such as multimodal large language models (MLLMs) [1, 5, 44] can yield more flexible and interactive geographic textual information but may not offer precise administrative addresses. In future work, we plan to explore the potential of MLLMs for the IAL task. To this end, we have made an attempt by constructing a question-and-answer dataset based on Pitts-IAL and adopting the LLaVA-1.5 [31] model for instruct tuning. Tab. 7 shows some examples comparing with representative MLLMs. It is evident that LLaVA-IAL, fine-tuned with instructions, can predict accurate administrative addresses consistent with address hierarchy, while other compared MLLMs are unable to predict addresses without landmarks or street signs and lack standardized output formats.
+
+# 7 Conclusion
+
+In this study, we introduce the problem of image address localization and propose three IAL datasets for evaluation and subsequent research. To facilitate the alignment of images and addresses for tackling the problem, we propose the AddressCLIP framework consisting of image-text alignment and image-geography matching. Extensive experiments on the proposed datasets validate that our method outperforms transfer learning methods that transfer CLIP to downstream tasks. We compare the proposed method with the existing two-stage address localization pipeline based on the image geo-localization technology and discuss AddressCLIP's application in real-world situations. Finally, we explore the potential of multimodal large language models for address localization.
+
+# Acknowledgements
+
+This work was supported by the National Natural Science Foundations of China (Grants No.62376267, 62076242) and the innoHK project.
+
+# References
+
+1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)
+2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022)
+3. Ali-Bey, A., Chaib-Draa, B., Giguere, P.: Mixvpr: Feature mixing for visual place recognition. In: WACV. pp. 2998-3007 (2023)
+4. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: CVPR. pp. 5297-5307 (2016)
+5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)
+6. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer vision and image understanding 110(3), 346-359 (2008)
+7. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: CVPR. pp. 4878-4888 (2022)
+8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023)
+9. Berton, G.M., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: WACV. pp. 2918-2927 (2021)
+0. Cepeda, V.V., Nayak, G.K., Shah, M.: Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. arXiv preprint arXiv:2309.16020 (2023)
+1. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)
+2. Clark, B., Kerrigan, A., Kulkarni, P.P., Cepeda, V.V., Shah, M.: Where we are and what we're looking at: Query based worldwide image geo-localization using hierarchies and scenes. In: CVPR. pp. 23182-23190 (2023)
+3. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: ECCV Workshop. vol. 1, pp. 1-2 (2004)
+4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)
+5. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)
+6. Ge, Y., Wang, H., Zhu, F., Zhao, R., Li, H.: Self-supervising fine-grained region similarities for large-scale image localization. In: ECCV. pp. 369-386 (2020)
+
+17. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: End-to-end learning of deep visual representations for image retrieval. International Journal of Computer Vision 124(2), 237-254 (2017)
+18. Haas, L., Alberti, S., Skreta, M.: Learning generalized zero-shot learners for open-domain image geolocation. arXiv preprint arXiv:2302.00275 (2023)
+19. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: CVPR. pp. 14141-14152 (2021)
+20. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020)
+21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)
+22. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)
+23. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)
+24. Jegou, H., Douze, M., Schmid, C.: Hamming embedding and weak geometric consistency for large scale image search. In: ECCV. pp. 304-317 (2008)
+25. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE transactions on pattern analysis and machine intelligence 34(9), 1704-1716 (2011)
+26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)
+27. Jin Kim, H., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geo-localization. In: CVPR. pp. 2136-2145 (2017)
+28. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)
+29. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR. pp. 19113-19122 (2023)
+30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900 (2022)
+31. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)
+32. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: ICCV. pp. 2570-2579 (2019)
+33. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)
+34. Oliva, A., Torralba, A.: Building the gist of a scene: The role of global image features in recognition. Progress in brain research 155, 23-36 (2006)
+35. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)
+36. Pramanick, S., Nowara, E.M., Gleason, J., Castillo, C.D., Chellappa, R.: Where in the world is this image? transformer-based geo-localization in the wild. In: ECCV. pp. 196-215 (2022)
+37. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE transactions on pattern analysis and machine intelligence 41(7), 1655-1668 (2018)
+
+38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)
+39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)
+40. Schindler, G., Brown, M., Szeliski, R.: City-scale location recognition. In: CVPR. pp. 1-7. IEEE (2007)
+41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv abs/2210.08402 (2022), https://api_semanticscholar.org/CorpusID: 252917726
+42. Seo, P.H., Weyand, T., Sim, J., Han, B.: Cplanet: Enhancing image geolocation by combinatorial partitioning of maps. In: ECCV. pp. 536-551 (2018)
+43. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)
+44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)
+45. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: CVPR. pp. 883-890 (2013)
+46. Trivigno, G., Berton, G., Aragon, J., Caputo, B., Masone, C.: Divide&classify: Fine-grained classification for city-wide visual geo-localization. In: ICCV. pp. 11142-11152 (2023)
+47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)
+48. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: CVPR. pp. 13648-13657 (2022)
+49. Weyand, T., Kostrikov, I., Philbin, J.: Planet-photo geolocation with convolutional neural networks. In: ECCV. pp. 37-55 (2016)
+50. Wilson, D., Zhang, X., Sultani, W., Wshah, S.: Image and object geo-localization. International Journal of Computer Vision pp. 1-43 (2023)
+51. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022)
+52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)
+53. Zhu, S., Yang, L., Chen, C., Shah, M., Shen, X., Wang, H.: R2former: Unified retrieval and reranking transformer for place recognition. In: CVPR. pp. 19370-19380 (2023)
+54. Zhu, Y., Wang, J., Xie, L., Zheng, L.: Attention-based pyramid aggregation network for visual place recognition. In: ACM MM. pp. 99-107 (2018)
\ No newline at end of file
diff --git a/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/images.zip b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/images.zip
new file mode 100644
index 0000000000000000000000000000000000000000..1fd3781a1c19a45754a0b72880a1e5fa744c3f37
--- /dev/null
+++ b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/images.zip
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dd351c2ed0d6d4a1df4a298b94210dccb1ccc9ad8626b062bea89feec779c3e
+size 652244
diff --git a/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/layout.json b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/layout.json
new file mode 100644
index 0000000000000000000000000000000000000000..4170b05affcb4f9e7e113f84e2b2a0261a076d2a
--- /dev/null
+++ b/2024/AddressCLIP_ Empowering Vision-Language Models for City-wide Image Address Localization/layout.json
@@ -0,0 +1,9775 @@
+{
+ "pdf_info": [
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 134,
+ 111,
+ 479,
+ 148
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 134,
+ 111,
+ 479,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 134,
+ 111,
+ 479,
+ 148
+ ],
+ "type": "text",
+ "content": "AddressCLIP: Empowering Vision-Language Models for City-wide Image Address Localization"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "text",
+ "content": "Shixiong "
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "\\mathrm{Xu}^{1,3*†}"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "text",
+ "content": ", Chenghao Zhang"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "^{2*}"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "text",
+ "content": ", Lubin Fan"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "^{2‡}"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "text",
+ "content": ", Gaofeng Meng"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "^{1,3,4‡}"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "text",
+ "content": ", Shiming Xiang"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "^{1,3}"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "text",
+ "content": ", and Jieping Ye"
+ },
+ {
+ "bbox": [
+ 133,
+ 167,
+ 479,
+ 194
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 175,
+ 201,
+ 436,
+ 223
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 175,
+ 201,
+ 436,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 175,
+ 201,
+ 436,
+ 223
+ ],
+ "type": "inline_equation",
+ "content": "^{1}"
+ },
+ {
+ "bbox": [
+ 175,
+ 201,
+ 436,
+ 223
+ ],
+ "type": "text",
+ "content": " MAIS, Institute of Automation, Chinese Academy of Sciences "
+ },
+ {
+ "bbox": [
+ 175,
+ 201,
+ 436,
+ 223
+ ],
+ "type": "inline_equation",
+ "content": "^{2}"
+ },
+ {
+ "bbox": [
+ 175,
+ 201,
+ 436,
+ 223
+ ],
+ "type": "text",
+ "content": " Alibaba Cloud"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 148,
+ 224,
+ 463,
+ 246
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 5,
+ "blocks": [
+ {
+ "bbox": [
+ 148,
+ 224,
+ 463,
+ 236
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 148,
+ 224,
+ 463,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 148,
+ 224,
+ 463,
+ 236
+ ],
+ "type": "inline_equation",
+ "content": "^{3}"
+ },
+ {
+ "bbox": [
+ 148,
+ 224,
+ 463,
+ 236
+ ],
+ "type": "text",
+ "content": " School of Artificial Intelligence, University of Chinese Academy of Sciences"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 148,
+ 235,
+ 463,
+ 246
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 148,
+ 235,
+ 463,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 148,
+ 235,
+ 463,
+ 246
+ ],
+ "type": "inline_equation",
+ "content": "^{4}"
+ },
+ {
+ "bbox": [
+ 148,
+ 235,
+ 463,
+ 246
+ ],
+ "type": "text",
+ "content": " CAIR, HK Institute of Science & Innovation, Chinese Academy of Sciences"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 281,
+ 455,
+ 479
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 159,
+ 281,
+ 455,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 281,
+ 455,
+ 479
+ ],
+ "type": "text",
+ "content": "Abstract. In this study, we introduce a new problem raised by social media and photojournalism, named Image Address Localization (IAL), which aims to predict the readable textual address where an image was taken. Existing two-stage approaches involve predicting geographical coordinates and converting them into human-readable addresses, which can lead to ambiguity and be resource-intensive. In contrast, we propose an end-to-end framework named AddressCLIP to solve the problem with more semantics, consisting of two key ingredients: i) image-text alignment to align images with addresses and scene captions by contrastive learning, and ii) image-geography matching to constrain image features with the spatial distance in terms of manifold learning. Additionally, we have built three datasets from Pittsburgh and San Francisco on different scales specifically for the IAL problem. Experiments demonstrate that our approach achieves compelling performance on the proposed datasets and outperforms representative transfer learning methods for vision-language models. Furthermore, extensive ablations and visualizations exhibit the effectiveness of the proposed method. The datasets and source code are available at https://github.com/xsx1001/AddressCLIP."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "text",
+ "content": "Keywords: Image address localization "
+ },
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "inline_equation",
+ "content": "\\cdot"
+ },
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "text",
+ "content": " Image-text alignment "
+ },
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "inline_equation",
+ "content": "\\cdot"
+ },
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "text",
+ "content": " Image-geography matching "
+ },
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "inline_equation",
+ "content": "\\cdot"
+ },
+ {
+ "bbox": [
+ 159,
+ 489,
+ 453,
+ 512
+ ],
+ "type": "text",
+ "content": " Vision-language model"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 132,
+ 534,
+ 230,
+ 547
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 534,
+ 230,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 534,
+ 230,
+ 547
+ ],
+ "type": "text",
+ "content": "1 Introduction"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 130,
+ 562,
+ 482,
+ 623
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 562,
+ 482,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 562,
+ 482,
+ 623
+ ],
+ "type": "text",
+ "content": "Users on social media platforms such as Facebook and Instagram often tag their pictures with textual addresses to connect with local communities, raising the demand for predicting the descriptive address information of the place where an image was taken. This has various practical applications, for instance, businesses and travel platforms can use addresses of images to provide recommendations"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 141,
+ 631,
+ 441,
+ 643
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 631,
+ 441,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 631,
+ 441,
+ 643
+ ],
+ "type": "text",
+ "content": "This work was done when Shixiong Xu was an intern at Alibaba Cloud."
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 143,
+ 643,
+ 232,
+ 653
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 643,
+ 232,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 643,
+ 232,
+ 653
+ ],
+ "type": "text",
+ "content": "* Equal contributions"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 143,
+ 653,
+ 244,
+ 665
+ ],
+ "type": "page_footnote",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 143,
+ 653,
+ 244,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 143,
+ 653,
+ 244,
+ 665
+ ],
+ "type": "text",
+ "content": "‡ Corresponding authors"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 0
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 133,
+ 114,
+ 294,
+ 235
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 294,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 294,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 294,
+ 235
+ ],
+ "type": "image",
+ "image_path": "695334a9387409baafc688a6a632ef6d942c7adedd72b49428d551a1211318b8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 130,
+ 239,
+ 482,
+ 275
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 239,
+ 482,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 239,
+ 482,
+ 275
+ ],
+ "type": "text",
+ "content": "Fig. 1: Comparison of image-based geo-localization and address localization tasks. The objective of the proposed task is to predict the semantic text address of a given image instead of a digital GPS coordinate without the need for a retrieval gallery."
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 300,
+ 114,
+ 482,
+ 235
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 300,
+ 114,
+ 482,
+ 235
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 300,
+ 114,
+ 482,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 300,
+ 114,
+ 482,
+ 235
+ ],
+ "type": "image",
+ "image_path": "d5b1f7bf1f3f3617bc00703413aa8efdd7baf6c5dd6a42f86b1245effd5a6d4b.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 130,
+ 300,
+ 479,
+ 324
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 300,
+ 479,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 300,
+ 479,
+ 324
+ ],
+ "type": "text",
+ "content": "or organize location-specific content. Additionally, photojournalism can rapidly verify the authenticity of the event with the image's address."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 130,
+ 327,
+ 482,
+ 458
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 327,
+ 482,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 327,
+ 482,
+ 458
+ ],
+ "type": "text",
+ "content": "To predict an image's address, one reasonable approach involves leveraging image geo-localization technology to predict GPS coordinates (i.e., latitude and longitude) from an image [50], followed by the reverse Geocoding to query for a readable address. Image geo-localization, also known as visual place recognition, is commonly treated as an image retrieval problem where a database of geo-tagged images serves as a matching reference for the query image. Previous retrieval-based methods [3,4,7,19,28] have shown remarkable performance. However, in practice, the creation of pre-collected geo-tagged databases requires significant labor and storage resources, while GPS coordinates lack readability and semantics. In addition, the conversion from GPS to readable addresses often presents ambiguities, and the Image-GPS-Address pipeline is not end-to-end."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 130,
+ 460,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 460,
+ 482,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 460,
+ 482,
+ 604
+ ],
+ "type": "text",
+ "content": "To alleviate the above issues, in this study, we propose to perform Image Address Localization (IAL) where a model is tasked to predict the readable textual address where a given image was taken. We design a semantic address partition strategy to perform fine-grained partitioning of city-wide addresses, conforming to the way humans describe address information. By doing this, we are able to train models in an end-to-end manner, and during inference, there is no need to construct a retrieval database which greatly reduces the storage and retrieval burden. Furthermore, the model's output addresses align more closely with human description habits, which provides a bridge for subsequent city-wide scene understanding and point-of-interest recommendation. Fig. 1 shows the comparison of image geo-localization and image address localization tasks, where the latter focuses on predicting human-readable textual address information."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 130,
+ 605,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 605,
+ 482,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 605,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "content": "In this study, we propose an end-to-end framework, AddressCLIP, based on the visual-language model CLIP [38], aiming to learn an alignment of images and addresses. Our approach leverages two key ingredients: image-text alignment and image-geography matching. Firstly, we introduce additional scene captions as a supplement to address text thus facilitating the alignment of images and tex"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "text",
+ "content": "2"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "text",
+ "content": "S. Xu et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 1
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 479,
+ 152
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 479,
+ 152
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 479,
+ 152
+ ],
+ "type": "text",
+ "content": "tual addresses by contrastive learning. Secondly, we propose an image-geography matching mechanism to bring features of geographically proximate images closer while separating features of images that are far apart geographically."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 130,
+ 152,
+ 481,
+ 247
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 152,
+ 481,
+ 247
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 152,
+ 481,
+ 247
+ ],
+ "type": "text",
+ "content": "To support the image address localization task, we constructed three IAL datasets of different sizes based on the Pitts-250k [4] and SF-XL datasets [7]: Pitts-IAL (234K), SF-IAL-Base (184K), and SF-IAL-Large (1.96M). In contrast to the original datasets, each image in our dataset is accompanied by not only its geographical coordinate but also the administrative address. Specifically, we utilized the reverse Geocoding API of Google Maps to retrieve administrative addresses for a portion of the images and obtain addresses for the remaining images through nearest-neighbor interpolation of the geographical coordinates."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "text",
+ "content": "We evaluate the proposed AddressCLIP framework on the introduced datasets. Our proposed method achieves a Top-1 address localization accuracy of over "
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "inline_equation",
+ "content": "80\\%"
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "text",
+ "content": " across three IAL datasets, most notably reaching a performance of "
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "inline_equation",
+ "content": "85.92\\%"
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "text",
+ "content": " on the largest dataset, SF-IAL-Large. Compared with challenging baselines [29, 51, 52] that transfer CLIP to the downstream IAL task, our AddressCLIP achieves improvements of "
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "inline_equation",
+ "content": "3\\%"
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "text",
+ "content": " to "
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "inline_equation",
+ "content": "6\\%"
+ },
+ {
+ "bbox": [
+ 130,
+ 247,
+ 482,
+ 367
+ ],
+ "type": "text",
+ "content": " on the proposed datasets. In addition, the qualitative results demonstrate good alignment between images and textual address queries in geographical space. Finally, we discuss the superiority of the proposed method over the two-stage \"Image-GPS-Address\" approach and explore the application prospects of multimodal large language models in the IAL task."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 146,
+ 368,
+ 347,
+ 379
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 368,
+ 347,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 368,
+ 347,
+ 379
+ ],
+ "type": "text",
+ "content": "Our contributions are summarized as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 137,
+ 387,
+ 479,
+ 518
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 10,
+ "blocks": [
+ {
+ "bbox": [
+ 138,
+ 387,
+ 479,
+ 422
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 387,
+ 479,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 387,
+ 479,
+ 422
+ ],
+ "type": "text",
+ "content": "- We formulate the image address localization problem and introduce the AddressCLIP framework for this problem by utilizing the alignment between the image and address text."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 138,
+ 423,
+ 479,
+ 458
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 423,
+ 479,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 423,
+ 479,
+ 458
+ ],
+ "type": "text",
+ "content": "- Two key ingredients are designed for better alignment of the image and address, i.e., image-caption alignment and image-geography matching, which are mutually beneficial."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 137,
+ 459,
+ 479,
+ 482
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 137,
+ 459,
+ 479,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 137,
+ 459,
+ 479,
+ 482
+ ],
+ "type": "text",
+ "content": "- We introduce three datasets named Pitts-IAL, SF-IAL-Base, and SF-IAL-Large to facilitate the study of the image address localization problem."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 137,
+ 483,
+ 479,
+ 518
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 137,
+ 483,
+ 479,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 137,
+ 483,
+ 479,
+ 518
+ ],
+ "type": "text",
+ "content": "- Experiments demonstrate that our method achieves compelling performance on the proposed IAL datasets. Extensive ablations, visualizations, and analyses are provided to show the effectiveness of the proposed method."
+ }
+ ]
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "sub_type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 544,
+ 237,
+ 558
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 544,
+ 237,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 544,
+ 237,
+ 558
+ ],
+ "type": "text",
+ "content": "2 Related Work"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 130,
+ 570,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 570,
+ 482,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 570,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "content": "Image Geo-localization. Image geo-localization, or visual place recognition, is usually formulated as an image retrieval problem on the city scale, which needs to collect a geo-tagged database of pre-computed embeddings of either local or global features [6,13,24,25,33,34,40,45]. In recent years, deep learning models [14, 21, 43] have been proven to perform remarkably in image feature extraction, complemented with an aggregation or pooling layer [3,4,9,16,17,27,32,37,54]. Recent methods achieve impressive retrieval performance by performing an additional re-ranking phase [19,48,53], adopting powerful pre-trained backbones [35]"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 392,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 392,
+ 91,
+ 447,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 392,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "text",
+ "content": "AddressCLIP"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 474,
+ 92,
+ 480,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 480,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 480,
+ 100
+ ],
+ "type": "text",
+ "content": "3"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 2
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 225,
+ 123
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 225,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 225,
+ 123
+ ],
+ "type": "text",
+ "content": "(a) Administrative Address"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 132,
+ 125,
+ 471,
+ 145
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 125,
+ 471,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 125,
+ 471,
+ 145
+ ],
+ "type": "text",
+ "content": "Example: Carnegie Mellon University, 5000, Forbes Avenue, Oakland, Pittsburgh, Pennsylvania, USA \nHierarchy: Building Name, House Number, Street, Neighborhood, City, County/State, Country"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 144,
+ 163,
+ 291,
+ 202
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 133,
+ 153,
+ 238,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 153,
+ 238,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 153,
+ 238,
+ 163
+ ],
+ "type": "text",
+ "content": "(b) Semantic Address Partition"
+ }
+ ]
+ }
+ ],
+ "index": 4,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 144,
+ 163,
+ 291,
+ 202
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 144,
+ 163,
+ 291,
+ 202
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 163,
+ 291,
+ 202
+ ],
+ "type": "image",
+ "image_path": "2f73cc3742b580501bb6c82b99cf31c2f92cccc6e46d64bef4870330eba688f6.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 5,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 149,
+ 206,
+ 282,
+ 229
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 149,
+ 206,
+ 282,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 149,
+ 206,
+ 282,
+ 229
+ ],
+ "type": "text",
+ "content": "located in South Side Flats, specifically on a section of East Carson Street lying between South 10th Street and South 12th Street"
+ }
+ ]
+ }
+ ],
+ "index": 6,
+ "angle": 0,
+ "type": "image_footnote"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 309,
+ 163,
+ 478,
+ 219
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 307,
+ 153,
+ 399,
+ 163
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 307,
+ 153,
+ 399,
+ 163
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 307,
+ 153,
+ 399,
+ 163
+ ],
+ "type": "text",
+ "content": "(c) Vision-Language Model"
+ }
+ ]
+ }
+ ],
+ "index": 7,
+ "angle": 0,
+ "type": "image_caption"
+ },
+ {
+ "bbox": [
+ 309,
+ 163,
+ 478,
+ 219
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 163,
+ 478,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 163,
+ 478,
+ 219
+ ],
+ "type": "image",
+ "image_path": "2dbe6bac1986642ec50383f38a41f704b56ed23319cd1b21aaad5641b3143f1e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 8,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 311,
+ 220,
+ 375,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 311,
+ 220,
+ 375,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 220,
+ 375,
+ 228
+ ],
+ "type": "text",
+ "content": "Candidate Addresses"
+ }
+ ]
+ }
+ ],
+ "index": 9,
+ "angle": 0,
+ "type": "image_footnote"
+ },
+ {
+ "bbox": [
+ 387,
+ 217,
+ 476,
+ 228
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 387,
+ 217,
+ 476,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 387,
+ 217,
+ 476,
+ 228
+ ],
+ "type": "text",
+ "content": "located in the Downtown, specifically on a section of Third Avenue that near Grant Street"
+ }
+ ]
+ }
+ ],
+ "index": 10,
+ "angle": 0,
+ "type": "image_footnote"
+ },
+ {
+ "bbox": [
+ 130,
+ 237,
+ 480,
+ 270
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 237,
+ 480,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 237,
+ 480,
+ 270
+ ],
+ "type": "text",
+ "content": "Fig. 2: The problem statement of the image address localization task consists of (a) examples of administrative address and hierarchy, (b) semantic address partition strategy, and (c) address predicting using visual-language models."
+ }
+ ]
+ }
+ ],
+ "index": 11,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 130,
+ 302,
+ 482,
+ 434
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 302,
+ 482,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 302,
+ 482,
+ 434
+ ],
+ "type": "text",
+ "content": "to extract image features [22, 28], or training on large-scale place recognition datasets [2, 3, 7, 22]. Different from retrieval-based methods, classification-based methods focus on planet-scale localization and split the earth into disjoint regions to classify [12, 36, 42, 46, 49]. More recently, StreetCLIP [18] and GeoCLIP [10] both utilize the vision-language model CLIP [38] with region description or GPS information for better generalizability. Going beyond image geo-localization, we propose to perform image address localization to obtain readable textual addresses rather than digital coordinates without a retrieval gallery. This not only enables models to directly output human-understandable semantic addresses for a given image but also paves the way for more complex geographical human-computer interactions in the future."
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 130,
+ 437,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 437,
+ 482,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 437,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "content": "Transfer Learning in Vision-Language Models. The integration of language supervision with visual data is garnering significant interest, with the primary aim being to align images and texts and learn a shared embedding space. As outlined in [52], the advancements in vision-language models can largely be attributed to three key developments: Transformers [47], contrastive representation learning [11, 20], and expansive web-scale training datasets [26, 39]. One notable example is CLIP [38], which employs two encoder networks trained via contrastive loss to align image-text pairs, thus enabling impressive zero-shot performance. Adapting CLIP to downstream tasks typically involves either full fine-tuning or linear probing [15]. Recently, prompt learning offers an alternative by introducing a small number of trainable prompt tokens at the input. Learnable prompts can be applied to the language branch [52], image instances [51], or both forming a multi-modal prompt [29]. Complete fine-tuning enables CLIP to fully adapt to the data distribution of downstream tasks, while prompt learning enhances CLIP's zero-shot learning capabilities. Due to the domain gap between the IAL task and the pre-training tasks, our proposed AddressCLIP adopts carefully designed image-caption alignment and image-geography matching to transfer CLIP toward the address localization task, which is superior to the direct complete fine-tuning way."
+ }
+ ]
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "text",
+ "content": "4"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "text",
+ "content": "S. Xu et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 3
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 132,
+ 114,
+ 270,
+ 127
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 114,
+ 270,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 114,
+ 270,
+ 127
+ ],
+ "type": "text",
+ "content": "3 Problem Statement"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 130,
+ 138,
+ 482,
+ 222
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 138,
+ 482,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 138,
+ 482,
+ 222
+ ],
+ "type": "text",
+ "content": "In this study, we focus on the city-wide image address localization problem. The administrative address hierarchy around the world varies widely depending on the history, geography, culture, and political systems of each country. Taking the United States as an example, we provide a specific illustration of an administrative address and its corresponding hierarchy in Fig. 2 (a). Since images in one dataset belong to the same city, our study distinguishes image addresses on neighborhood and street levels."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 130,
+ 223,
+ 482,
+ 390
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 223,
+ 482,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 223,
+ 482,
+ 390
+ ],
+ "type": "text",
+ "content": "The straightforward division mentioned above introduces two challenges in practical city-wide scenarios. Firstly, variable street lengths can result in coarsely localized addresses, particularly for highways that extend for kilometers, creating a pronounced long-tail distribution issue and diverse inner-address visual features that hinder precise localization during inference. Secondly, address ambiguity arises at street intersections, where images could be equally attributed to intersecting streets, thus lacking a clear and singular textual supervision signal. To address these concerns, we introduce a semantic address partition strategy for a more granular segmentation of streets, as shown in Fig. 2 (b). By segmenting streets at intersections, we achieve a balance in street lengths, which refines the address localization scope and eliminates the intersection ambiguity, aligning more closely with the way humans typically describe locations. In this way, the textual representation of addresses consists of the main street name (marked green) and the name of one or two streets that intersect it (marked brown)."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": "Formally, the Image Address Localization problem is defined as follows: given a training dataset "
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "D_{train} = \\{(I_i,A_i)\\}_{i = 1}^M"
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": " containing pairs of image "
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "I_{i}"
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": " and address "
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "A_{i}"
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": ", our objective is to train a vision-language model "
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{H}_{\\theta}"
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": " and use it to predict the address of query images, "
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "A_{k}^{Q} = \\mathcal{H}_{\\theta}(I_{k}^{Q}),\\forall k\\in [1..K]"
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": " where "
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "I_k^Q\\in D_{test}"
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": ". The images in the query set "
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "inline_equation",
+ "content": "I^Q"
+ },
+ {
+ "bbox": [
+ 130,
+ 390,
+ 482,
+ 475
+ ],
+ "type": "text",
+ "content": " can belong to any candidate address in the same city as the images in the training set. Fig. 2 (c) shows a schematic diagram of predicting the readable textual address for a given query image."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 132,
+ 491,
+ 234,
+ 503
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 491,
+ 234,
+ 503
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 491,
+ 234,
+ 503
+ ],
+ "type": "text",
+ "content": "4 AddressCLIP"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 132,
+ 515,
+ 268,
+ 526
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 515,
+ 268,
+ 526
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 515,
+ 268,
+ 526
+ ],
+ "type": "text",
+ "content": "4.1 Framework Overview"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 533,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "content": "We formulate the IAL problem as a vision-text alignment problem between the image and address pairs. Fig. 3 depicts the framework of our method. During training, the embeddings of the image and the address are extracted by the image encoder and text encoder, respectively, and are then aligned through image-address contrastive learning. An additive scene caption is introduced as a supplement to the address to enrich the plain text information. The scene caption shares the same text encoder with the image address, and the resulting caption embedding and image embedding are combined for image-caption contrastive learning. Furthermore, we adopt the geographical position information as a guide to increase the similarities between geographically close image features while increasing the differences between geographically distant image features."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 101
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 101
+ ],
+ "type": "text",
+ "content": "AddressCLIP"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 474,
+ 91,
+ 481,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 474,
+ 91,
+ 481,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 474,
+ 91,
+ 481,
+ 100
+ ],
+ "type": "text",
+ "content": "5"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 4
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 133,
+ 114,
+ 482,
+ 278
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 482,
+ 278
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 482,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 482,
+ 278
+ ],
+ "type": "image",
+ "image_path": "8cd9089a8a9eb31a36c474a70144488dd181e20df4e580f902dd13d846b0393f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "image_body"
+ },
+ {
+ "bbox": [
+ 130,
+ 285,
+ 482,
+ 330
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 285,
+ 482,
+ 330
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 285,
+ 482,
+ 330
+ ],
+ "type": "text",
+ "content": "Fig. 3: Overview of the proposed AddressCLIP framework. (a) During training, the alignment of image and address is learned by the image-address contrastive loss, image-capture contrastive loss, and image-geography matching loss. (b) At inference, the address with the highest similarity to the query image's embedding is chosen."
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "image_caption"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 130,
+ 355,
+ 482,
+ 392
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 355,
+ 482,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 355,
+ 482,
+ 392
+ ],
+ "type": "text",
+ "content": "The image-geography matching is learned between geospatial distance and image feature similarity. During inference, the address with the highest similarity to the query image's embedding indicates the most probable address."
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 132,
+ 413,
+ 275,
+ 426
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 413,
+ 275,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 413,
+ 275,
+ 426
+ ],
+ "type": "text",
+ "content": "4.2 Image-Text Alignment"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 130,
+ 437,
+ 482,
+ 544
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 437,
+ 482,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 437,
+ 482,
+ 544
+ ],
+ "type": "text",
+ "content": "It is reasonable to use address information directly as textual prompts for image-address alignment learning. However, the address text is simple and limited. It cannot provide context about environments, landmarks, or other entities, which are crucial for precise address localization. To alleviate the issues, we incorporate additional descriptive captions that capture the nuances of the visual scene, thereby endowing the model with a deeper understanding of the contextual elements that are often missing in the bare address labels. This mechanism enables more accurate and context-aware predictions by effectively bridging the gap between visual perception and textual representation."
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 130,
+ 546,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 546,
+ 482,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 546,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "content": "Scene description can be generated through manual annotation, which, although accurate, is costly and not easily scalable to large datasets. Benefiting from the advancements in vision-language models, we utilize pre-trained vision-language models [30] to generate linguistic captions corresponding to image scenes. The lower left corner of Fig. 3(a) shows some illustrative examples, where the descriptions can include context like the presence of specific buildings or unique street signs, which is relevant for distinguishing between visually similar but geographically distant locations. This also aligns the model's learning process with how humans typically communicate location information. For detailed analyses of scene captions, refer to the supplementary material."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 140,
+ 100
+ ],
+ "type": "text",
+ "content": "6"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "text",
+ "content": "S. Xu et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 5
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "content": "Formally, define image features extracted from the image encoder "
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{V}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "content": " as "
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "inline_equation",
+ "content": "V_{i} = \\mathcal{V}(I_{i}),\\forall i\\in [1,\\dots,N]"
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "content": ". The text encoder "
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "inline_equation",
+ "content": "\\mathcal{V}(\\cdot)"
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "content": " extracts address features "
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "inline_equation",
+ "content": "T_{i}^{A} = \\mathcal{T}(A_{i})"
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "content": " and caption features "
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "inline_equation",
+ "content": "T_{i}^{C} = \\mathcal{T}(C_{i} + A_{i})"
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "content": ", where the scene caption "
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "inline_equation",
+ "content": "C_i"
+ },
+ {
+ "bbox": [
+ 131,
+ 116,
+ 482,
+ 212
+ ],
+ "type": "text",
+ "content": " is obtained by a vision-language model. We experimentally observe that appending address information to the scene caption is more conducive to address localization, which is discussed in detail in Sec. 6.3. Note that the additive scene caption is only used for training. The alignment of images and addresses is learned via image-address contrastive loss and image-caption contrastive loss."
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 131,
+ 212,
+ 482,
+ 236
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 212,
+ 482,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 212,
+ 482,
+ 236
+ ],
+ "type": "text",
+ "content": "For a batch of size "
+ },
+ {
+ "bbox": [
+ 131,
+ 212,
+ 482,
+ 236
+ ],
+ "type": "inline_equation",
+ "content": "N"
+ },
+ {
+ "bbox": [
+ 131,
+ 212,
+ 482,
+ 236
+ ],
+ "type": "text",
+ "content": " comprising image-text pairs, the image-address contrastive loss can be written as:"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 146,
+ 240,
+ 481,
+ 270
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 240,
+ 481,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 240,
+ 481,
+ 270
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} _ {\\text {a d d r e s s}} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left[ \\log \\frac {\\exp \\left(V _ {i} \\cdot T _ {i} ^ {A} / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(V _ {i} \\cdot T _ {j} ^ {A} / \\tau\\right)} + \\log \\frac {\\exp \\left(T _ {i} ^ {A} \\cdot V _ {i} / \\tau\\right)}{\\sum_ {k = 1} ^ {N} \\exp \\left(T _ {i} ^ {A} \\cdot V _ {k} / \\tau\\right)} \\right], \\tag {1}",
+ "image_path": "c16bfd72eb064167a12a953b601036b8d23a4a9bfbba759512b8fc7773c3cf3d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 131,
+ 275,
+ 482,
+ 299
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 275,
+ 482,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 275,
+ 482,
+ 299
+ ],
+ "type": "text",
+ "content": "where "
+ },
+ {
+ "bbox": [
+ 131,
+ 275,
+ 482,
+ 299
+ ],
+ "type": "inline_equation",
+ "content": "\\tau"
+ },
+ {
+ "bbox": [
+ 131,
+ 275,
+ 482,
+ 299
+ ],
+ "type": "text",
+ "content": " is the temperature parameter. Similarly, the image-caption contrastive loss is formulated as:"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 146,
+ 303,
+ 481,
+ 334
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 303,
+ 481,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 303,
+ 481,
+ 334
+ ],
+ "type": "interline_equation",
+ "content": "\\mathcal {L} _ {c a p t i o n} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left[ \\log \\frac {\\exp (V _ {i} \\cdot T _ {i} ^ {C} / \\tau)}{\\sum_ {j = 1} ^ {N} \\exp (V _ {i} \\cdot T _ {j} ^ {C} / \\tau)} + \\log \\frac {\\exp (T _ {i} ^ {C} \\cdot V _ {i} / \\tau)}{\\sum_ {k = 1} ^ {N} \\exp (T _ {i} ^ {C} \\cdot V _ {k} / \\tau)} \\right]. \\quad (2)",
+ "image_path": "f98d81d9444ec47ceb0044e65297df415080f0b2ddb7313a9376552c0af9234e.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 132,
+ 347,
+ 302,
+ 360
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 347,
+ 302,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 347,
+ 302,
+ 360
+ ],
+ "type": "text",
+ "content": "4.3 Image-Geography Matching"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 131,
+ 365,
+ 482,
+ 497
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 365,
+ 482,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 365,
+ 482,
+ 497
+ ],
+ "type": "text",
+ "content": "In general, address text in city-wide scenarios may be geographically far away but highly similar, or geographically close but significantly different. This makes image-address alignment learning difficult to optimize with address text alone. In contrast, the geographic coordinates of images (e.g., UTM coordinates) differ significantly, showcasing clear distinctions and discriminative properties. From the perspective of manifold learning, image embedding represents a low-dimensional representation of images in the feature space, and its distribution should be consistent with the geographic coordinates of the images. Our goal is to ensure that geographically proximate images exhibit closely in the feature space, while geographically distant images reflect more within the feature space. Visualization results and analysis are elaborated in Sec. 6.4."
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "text",
+ "content": "Inspired by the above motivation, we propose an image-geography matching loss to constrain image features according to the spatial distances of geographic coordinates. Specifically, denote "
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "inline_equation",
+ "content": "U_{i}:\\mathbb{UTM}_{east}\\times \\mathbb{UTM}_{north},\\forall i\\in [1,\\dots ,N]"
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "text",
+ "content": " the set of geographic coordinates corresponding to all images within a batch of size "
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "inline_equation",
+ "content": "N"
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "text",
+ "content": ". We can calculate each element of the spatial distance matrix "
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "inline_equation",
+ "content": "D^{U}"
+ },
+ {
+ "bbox": [
+ 131,
+ 498,
+ 482,
+ 570
+ ],
+ "type": "text",
+ "content": " in the geographic space as follows:"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 200,
+ 576,
+ 481,
+ 602
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 200,
+ 576,
+ 481,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 200,
+ 576,
+ 481,
+ 602
+ ],
+ "type": "interline_equation",
+ "content": "D _ {i j} ^ {U} = \\left\\| \\hat {U} _ {i} - \\hat {U} _ {j} \\right\\| _ {1}, s. t., \\hat {U} _ {i} = \\frac {U _ {i} - \\operatorname* {m i n} \\left(U _ {i}\\right)}{\\operatorname* {m a x} \\left(U _ {i}\\right) - \\operatorname* {m i n} \\left(U _ {i}\\right)}, \\tag {3}",
+ "image_path": "51ca724c95c4b7826ff87b867ccd7582c0454cac9c8e1c41687fd7d0e7acb183.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 131,
+ 609,
+ 482,
+ 644
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 609,
+ 482,
+ 644
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 609,
+ 482,
+ 644
+ ],
+ "type": "text",
+ "content": "where Manhattan distance and min-max normalization are adopted. Correspondingly, each element of the feature similarity matrix "
+ },
+ {
+ "bbox": [
+ 131,
+ 609,
+ 482,
+ 644
+ ],
+ "type": "inline_equation",
+ "content": "D^{V}"
+ },
+ {
+ "bbox": [
+ 131,
+ 609,
+ 482,
+ 644
+ ],
+ "type": "text",
+ "content": " in the image embedding space is calculated as:"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 266,
+ 643,
+ 481,
+ 669
+ ],
+ "type": "interline_equation",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 266,
+ 643,
+ 481,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 266,
+ 643,
+ 481,
+ 669
+ ],
+ "type": "interline_equation",
+ "content": "D _ {i j} ^ {V} = \\frac {V _ {i} \\cdot V _ {j}}{| | V _ {i} | | \\cdot | | V _ {j} | |}. \\tag {4}",
+ "image_path": "f008f4d5ad296e45a9cede64f3634aa66293011796cc9a2b1225127beaefb4f3.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "text",
+ "content": "AddressCLIP"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 474,
+ 92,
+ 481,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 481,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 474,
+ 92,
+ 481,
+ 100
+ ],
+ "type": "text",
+ "content": "7"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 6
+ },
+ {
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 144,
+ 129,
+ 470,
+ 203
+ ],
+ "blocks": [
+ {
+ "bbox": [
+ 134,
+ 117,
+ 478,
+ 128
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 134,
+ 117,
+ 478,
+ 128
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 134,
+ 117,
+ 478,
+ 128
+ ],
+ "type": "text",
+ "content": "Table 1: Detailed information of the proposed Image Address Localization datasets."
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "angle": 0,
+ "type": "table_caption"
+ },
+ {
+ "bbox": [
+ 144,
+ 129,
+ 470,
+ 203
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 144,
+ 129,
+ 470,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 144,
+ 129,
+ 470,
+ 203
+ ],
+ "type": "table",
+ "html": "| Visual input examples for Image Address Localization: |
| Grant Street, Downtown | Fifth Avenue, Downtown |
| User | Where might this photo have been taken? Tell me its street level address. |
| LLaVA-IAL | The address of this photo might be Grant Street, Downtown, Pittsburgh, PA, USA. | The address of this photo might be Fifth Avenue, Downtown, Pittsburgh, PA, USA. |
| GPT-4V [1] | This photo was taken in Peavey Plaza in Minneapolis, Minnesota, USA, on the corner of Nicolson Mall and 11th Street. | This photo was taken at 400 Fifth Avenue in Pittsburgh, Pennsylvania. |
| Gemini-Pro-V [44] | This photo was taken on the Rose Fitzgerald Kennedy Greenway in Boston, Massachusetts. | The photo was taken on Fifth Avenue in Pittsburgh, Pennsylvania. |
| QWen-VL [5] | This photo appears to have been taken on a city street, possibly in a downtown area or commercial district. | This photo appears to have been taken on Fifth Avenue in Pittsburgh, Pennsylvania, USA. The street sign in the image confirms this location. |
",
+ "image_path": "f4bba665d5caf19a7a481f1744ced7f846b1962d2de1de5a17ce4ceb4f0bc9d8.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 3,
+ "angle": 0,
+ "type": "table_body"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 132,
+ 335,
+ 309,
+ 346
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 335,
+ 309,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 335,
+ 309,
+ 346
+ ],
+ "type": "text",
+ "content": "6.6 Limitations and Future Work"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 130,
+ 356,
+ 482,
+ 499
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 356,
+ 482,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 356,
+ 482,
+ 499
+ ],
+ "type": "text",
+ "content": "The proposed AddressCLIP can be regarded as a discriminative model, limited by the set of candidate addresses at inference. In contrast, generative models such as multimodal large language models (MLLMs) [1, 5, 44] can yield more flexible and interactive geographic textual information but may not offer precise administrative addresses. In future work, we plan to explore the potential of MLLMs for the IAL task. To this end, we have made an attempt by constructing a question-and-answer dataset based on Pitts-IAL and adopting the LLaVA-1.5 [31] model for instruct tuning. Tab. 7 shows some examples comparing with representative MLLMs. It is evident that LLaVA-IAL, fine-tuned with instructions, can predict accurate administrative addresses consistent with address hierarchy, while other compared MLLMs are unable to predict addresses without landmarks or street signs and lack standardized output formats."
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 132,
+ 519,
+ 220,
+ 532
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 519,
+ 220,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 519,
+ 220,
+ 532
+ ],
+ "type": "text",
+ "content": "7 Conclusion"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 130,
+ 545,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 545,
+ 482,
+ 666
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 545,
+ 482,
+ 666
+ ],
+ "type": "text",
+ "content": "In this study, we introduce the problem of image address localization and propose three IAL datasets for evaluation and subsequent research. To facilitate the alignment of images and addresses for tackling the problem, we propose the AddressCLIP framework consisting of image-text alignment and image-geography matching. Extensive experiments on the proposed datasets validate that our method outperforms transfer learning methods that transfer CLIP to downstream tasks. We compare the proposed method with the existing two-stage address localization pipeline based on the image geo-localization technology and discuss AddressCLIP's application in real-world situations. Finally, we explore the potential of multimodal large language models for address localization."
+ }
+ ]
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 144,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 144,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 144,
+ 100
+ ],
+ "type": "text",
+ "content": "14"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "text",
+ "content": "S. Xu et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 13
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 246,
+ 129
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 246,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 114,
+ 246,
+ 129
+ ],
+ "type": "text",
+ "content": "Acknowledgements"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 132,
+ 141,
+ 480,
+ 166
+ ],
+ "type": "text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 141,
+ 480,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 141,
+ 480,
+ 166
+ ],
+ "type": "text",
+ "content": "This work was supported by the National Natural Science Foundations of China (Grants No.62376267, 62076242) and the innoHK project."
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 133,
+ 185,
+ 197,
+ 198
+ ],
+ "type": "title",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 185,
+ 197,
+ 198
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 185,
+ 197,
+ 198
+ ],
+ "type": "text",
+ "content": "References"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 138,
+ 211,
+ 481,
+ 665
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 21,
+ "blocks": [
+ {
+ "bbox": [
+ 138,
+ 211,
+ 481,
+ 244
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 211,
+ 481,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 211,
+ 481,
+ 244
+ ],
+ "type": "text",
+ "content": "1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 138,
+ 246,
+ 481,
+ 266
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 246,
+ 481,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 246,
+ 481,
+ 266
+ ],
+ "type": "text",
+ "content": "2. Ali-bey, A., Chaib-draa, B., Giguère, P.: Gsv-cities: Toward appropriate supervised visual place recognition. Neurocomputing 513, 194-203 (2022)"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 138,
+ 268,
+ 481,
+ 289
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 268,
+ 481,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 268,
+ 481,
+ 289
+ ],
+ "type": "text",
+ "content": "3. Ali-Bey, A., Chaib-Draa, B., Giguere, P.: Mixvpr: Feature mixing for visual place recognition. In: WACV. pp. 2998-3007 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 138,
+ 290,
+ 481,
+ 311
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 290,
+ 481,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 290,
+ 481,
+ 311
+ ],
+ "type": "text",
+ "content": "4. Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn architecture for weakly supervised place recognition. In: CVPR. pp. 5297-5307 (2016)"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 138,
+ 312,
+ 481,
+ 344
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 312,
+ 481,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 312,
+ 481,
+ 344
+ ],
+ "type": "text",
+ "content": "5. Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 138,
+ 346,
+ 481,
+ 366
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 346,
+ 481,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 346,
+ 481,
+ 366
+ ],
+ "type": "text",
+ "content": "6. Bay, H., Ess, A., Tuytelaars, T., Van Gool, L.: Speeded-up robust features (surf). Computer vision and image understanding 110(3), 346-359 (2008)"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 138,
+ 368,
+ 481,
+ 388
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 368,
+ 481,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 368,
+ 481,
+ 388
+ ],
+ "type": "text",
+ "content": "7. Berton, G., Masone, C., Caputo, B.: Rethinking visual geo-localization for large-scale applications. In: CVPR. pp. 4878-4888 (2022)"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 138,
+ 390,
+ 481,
+ 432
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 390,
+ 481,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 390,
+ 481,
+ 432
+ ],
+ "type": "text",
+ "content": "8. Berton, G., Trivigno, G., Caputo, B., Masone, C.: Eigenplaces: Training viewpoint robust models for visual place recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 11080-11090 (October 2023)"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 138,
+ 434,
+ 481,
+ 455
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 434,
+ 481,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 434,
+ 481,
+ 455
+ ],
+ "type": "text",
+ "content": "9. Berton, G.M., Paolicelli, V., Masone, C., Caputo, B.: Adaptive-attentive geolocation from few queries: A hybrid approach. In: WACV. pp. 2918-2927 (2021)"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 138,
+ 456,
+ 481,
+ 487
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 456,
+ 481,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 456,
+ 481,
+ 487
+ ],
+ "type": "text",
+ "content": "0. Cepeda, V.V., Nayak, G.K., Shah, M.: Geoclip: Clip-inspired alignment between locations and images for effective worldwide geo-localization. arXiv preprint arXiv:2309.16020 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 138,
+ 489,
+ 481,
+ 510
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 489,
+ 481,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 489,
+ 481,
+ 510
+ ],
+ "type": "text",
+ "content": "1. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: ICML. pp. 1597-1607 (2020)"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 138,
+ 512,
+ 481,
+ 543
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 512,
+ 481,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 512,
+ 481,
+ 543
+ ],
+ "type": "text",
+ "content": "2. Clark, B., Kerrigan, A., Kulkarni, P.P., Cepeda, V.V., Shah, M.: Where we are and what we're looking at: Query based worldwide image geo-localization using hierarchies and scenes. In: CVPR. pp. 23182-23190 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 138,
+ 544,
+ 481,
+ 565
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 544,
+ 481,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 544,
+ 481,
+ 565
+ ],
+ "type": "text",
+ "content": "3. Csurka, G., Dance, C., Fan, L., Willamowski, J., Bray, C.: Visual categorization with bags of keypoints. In: ECCV Workshop. vol. 1, pp. 1-2 (2004)"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 138,
+ 567,
+ 481,
+ 609
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 567,
+ 481,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 567,
+ 481,
+ 609
+ ],
+ "type": "text",
+ "content": "4. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 138,
+ 611,
+ 481,
+ 643
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 611,
+ 481,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 611,
+ 481,
+ 643
+ ],
+ "type": "text",
+ "content": "5. Gao, P., Geng, S., Zhang, R., Ma, T., Fang, R., Zhang, Y., Li, H., Qiao, Y.: Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision pp. 1-15 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 138,
+ 644,
+ 481,
+ 665
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 138,
+ 644,
+ 481,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 138,
+ 644,
+ 481,
+ 665
+ ],
+ "type": "text",
+ "content": "6. Ge, Y., Wang, H., Zhu, F., Zhao, R., Li, H.: Self-supervising fine-grained region similarities for large-scale image localization. In: ECCV. pp. 369-386 (2020)"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "text",
+ "content": "AddressCLIP"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 470,
+ 92,
+ 481,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 470,
+ 92,
+ 481,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 470,
+ 92,
+ 481,
+ 100
+ ],
+ "type": "text",
+ "content": "15"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 14
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 132,
+ 117,
+ 480,
+ 665
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 23,
+ "blocks": [
+ {
+ "bbox": [
+ 133,
+ 117,
+ 480,
+ 149
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 117,
+ 480,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 117,
+ 480,
+ 149
+ ],
+ "type": "text",
+ "content": "17. Gordo, A., Almazan, J., Revaud, J., Larlus, D.: End-to-end learning of deep visual representations for image retrieval. International Journal of Computer Vision 124(2), 237-254 (2017)"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 133,
+ 150,
+ 480,
+ 171
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 150,
+ 480,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 150,
+ 480,
+ 171
+ ],
+ "type": "text",
+ "content": "18. Haas, L., Alberti, S., Skreta, M.: Learning generalized zero-shot learners for open-domain image geolocation. arXiv preprint arXiv:2302.00275 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 133,
+ 172,
+ 480,
+ 203
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 172,
+ 480,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 172,
+ 480,
+ 203
+ ],
+ "type": "text",
+ "content": "19. Hausler, S., Garg, S., Xu, M., Milford, M., Fischer, T.: Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In: CVPR. pp. 14141-14152 (2021)"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 132,
+ 204,
+ 480,
+ 224
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 204,
+ 480,
+ 224
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 204,
+ 480,
+ 224
+ ],
+ "type": "text",
+ "content": "20. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729-9738 (2020)"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 132,
+ 225,
+ 480,
+ 246
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 225,
+ 480,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 225,
+ 480,
+ 246
+ ],
+ "type": "text",
+ "content": "21. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770-778 (2016)"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 132,
+ 247,
+ 480,
+ 268
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 247,
+ 480,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 247,
+ 480,
+ 268
+ ],
+ "type": "text",
+ "content": "22. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 132,
+ 269,
+ 480,
+ 289
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 269,
+ 480,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 269,
+ 480,
+ 289
+ ],
+ "type": "text",
+ "content": "23. Izquierdo, S., Civera, J.: Optimal transport aggregation for visual place recognition. arXiv preprint arXiv:2311.15937 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 132,
+ 290,
+ 480,
+ 310
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 290,
+ 480,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 290,
+ 480,
+ 310
+ ],
+ "type": "text",
+ "content": "24. Jegou, H., Douze, M., Schmid, C.: Hamming embedding and weak geometric consistency for large scale image search. In: ECCV. pp. 304-317 (2008)"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 132,
+ 311,
+ 480,
+ 342
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 311,
+ 480,
+ 342
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 311,
+ 480,
+ 342
+ ],
+ "type": "text",
+ "content": "25. Jégou, H., Perronnin, F., Douze, M., Sánchez, J., Pérez, P., Schmid, C.: Aggregating local image descriptors into compact codes. IEEE transactions on pattern analysis and machine intelligence 34(9), 1704-1716 (2011)"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 132,
+ 343,
+ 480,
+ 374
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 343,
+ 480,
+ 374
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 343,
+ 480,
+ 374
+ ],
+ "type": "text",
+ "content": "26. Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML. pp. 4904-4916 (2021)"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 132,
+ 376,
+ 480,
+ 396
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 376,
+ 480,
+ 396
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 376,
+ 480,
+ 396
+ ],
+ "type": "text",
+ "content": "27. Jin Kim, H., Dunn, E., Frahm, J.M.: Learned contextual feature reweighting for image geo-localization. In: CVPR. pp. 2136-2145 (2017)"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 132,
+ 396,
+ 480,
+ 429
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 396,
+ 480,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 396,
+ 480,
+ 429
+ ],
+ "type": "text",
+ "content": "28. Keetha, N., Mishra, A., Karhade, J., Jatavallabhula, K.M., Scherer, S., Krishna, M., Garg, S.: Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 132,
+ 430,
+ 480,
+ 450
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 430,
+ 480,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 430,
+ 480,
+ 450
+ ],
+ "type": "text",
+ "content": "29. Khattak, M.U., Rasheed, H., Maaz, M., Khan, S., Khan, F.S.: Maple: Multi-modal prompt learning. In: CVPR. pp. 19113-19122 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 132,
+ 451,
+ 480,
+ 482
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 451,
+ 480,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 451,
+ 480,
+ 482
+ ],
+ "type": "text",
+ "content": "30. Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: ICML. pp. 12888-12900 (2022)"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 132,
+ 483,
+ 480,
+ 504
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 483,
+ 480,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 483,
+ 480,
+ 504
+ ],
+ "type": "text",
+ "content": "31. Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 132,
+ 505,
+ 480,
+ 525
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 505,
+ 480,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 505,
+ 480,
+ 525
+ ],
+ "type": "text",
+ "content": "32. Liu, L., Li, H., Dai, Y.: Stochastic attraction-repulsion embedding for large scale image localization. In: ICCV. pp. 2570-2579 (2019)"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 132,
+ 526,
+ 480,
+ 547
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 526,
+ 480,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 526,
+ 480,
+ 547
+ ],
+ "type": "text",
+ "content": "33. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. International journal of computer vision 60, 91-110 (2004)"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 132,
+ 548,
+ 480,
+ 568
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 548,
+ 480,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 548,
+ 480,
+ 568
+ ],
+ "type": "text",
+ "content": "34. Oliva, A., Torralba, A.: Building the gist of a scene: The role of global image features in recognition. Progress in brain research 155, 23-36 (2006)"
+ }
+ ]
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 132,
+ 569,
+ 480,
+ 600
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 569,
+ 480,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 569,
+ 480,
+ 600
+ ],
+ "type": "text",
+ "content": "35. Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 132,
+ 601,
+ 480,
+ 633
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 601,
+ 480,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 601,
+ 480,
+ 633
+ ],
+ "type": "text",
+ "content": "36. Pramanick, S., Nowara, E.M., Gleason, J., Castillo, C.D., Chellappa, R.: Where in the world is this image? transformer-based geo-localization in the wild. In: ECCV. pp. 196-215 (2022)"
+ }
+ ]
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 132,
+ 634,
+ 480,
+ 665
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 132,
+ 634,
+ 480,
+ 665
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 132,
+ 634,
+ 480,
+ 665
+ ],
+ "type": "text",
+ "content": "37. Radenović, F., Tobias, G., Chum, O.: Fine-tuning cnn image retrieval with no human annotation. IEEE transactions on pattern analysis and machine intelligence 41(7), 1655-1668 (2018)"
+ }
+ ]
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 144,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 144,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 133,
+ 91,
+ 144,
+ 100
+ ],
+ "type": "text",
+ "content": "16"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 166,
+ 91,
+ 214,
+ 101
+ ],
+ "type": "text",
+ "content": "S. Xu et al."
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 15
+ },
+ {
+ "para_blocks": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 482,
+ 599
+ ],
+ "type": "list",
+ "angle": 0,
+ "index": 19,
+ "blocks": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 482,
+ 149
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 482,
+ 149
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 116,
+ 482,
+ 149
+ ],
+ "type": "text",
+ "content": "38. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)"
+ }
+ ]
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 130,
+ 150,
+ 482,
+ 182
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 150,
+ 482,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 150,
+ 482,
+ 182
+ ],
+ "type": "text",
+ "content": "39. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML. pp. 8748-8763 (2021)"
+ }
+ ]
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 130,
+ 183,
+ 481,
+ 205
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 130,
+ 183,
+ 481,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 130,
+ 183,
+ 481,
+ 205
+ ],
+ "type": "text",
+ "content": "40. Schindler, G., Brown, M., Szeliski, R.: City-scale location recognition. In: CVPR. pp. 1-7. IEEE (2007)"
+ }
+ ]
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 131,
+ 205,
+ 481,
+ 269
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 205,
+ 481,
+ 269
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 205,
+ 481,
+ 269
+ ],
+ "type": "text",
+ "content": "41. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J.: Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv abs/2210.08402 (2022), https://api_semanticscholar.org/CorpusID: 252917726"
+ }
+ ]
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 131,
+ 270,
+ 481,
+ 293
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 270,
+ 481,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 270,
+ 481,
+ 293
+ ],
+ "type": "text",
+ "content": "42. Seo, P.H., Weyand, T., Sim, J., Han, B.: Cplanet: Enhancing image geolocation by combinatorial partitioning of maps. In: ECCV. pp. 536-551 (2018)"
+ }
+ ]
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 131,
+ 293,
+ 481,
+ 314
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 293,
+ 481,
+ 314
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 293,
+ 481,
+ 314
+ ],
+ "type": "text",
+ "content": "43. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)"
+ }
+ ]
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 131,
+ 315,
+ 481,
+ 346
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 315,
+ 481,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 315,
+ 481,
+ 346
+ ],
+ "type": "text",
+ "content": "44. Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 131,
+ 347,
+ 481,
+ 369
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 347,
+ 481,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 347,
+ 481,
+ 369
+ ],
+ "type": "text",
+ "content": "45. Torii, A., Sivic, J., Pajdla, T., Okutomi, M.: Visual place recognition with repetitive structures. In: CVPR. pp. 883-890 (2013)"
+ }
+ ]
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 131,
+ 369,
+ 481,
+ 402
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 369,
+ 481,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 369,
+ 481,
+ 402
+ ],
+ "type": "text",
+ "content": "46. Trivigno, G., Berton, G., Aragon, J., Caputo, B., Masone, C.: Divide&classify: Fine-grained classification for city-wide visual geo-localization. In: ICCV. pp. 11142-11152 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 131,
+ 402,
+ 481,
+ 424
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 402,
+ 481,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 402,
+ 481,
+ 424
+ ],
+ "type": "text",
+ "content": "47. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. NeurIPS 30 (2017)"
+ }
+ ]
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 131,
+ 425,
+ 481,
+ 456
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 425,
+ 481,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 425,
+ 481,
+ 456
+ ],
+ "type": "text",
+ "content": "48. Wang, R., Shen, Y., Zuo, W., Zhou, S., Zheng, N.: Transvpr: Transformer-based place recognition with multi-level attention aggregation. In: CVPR. pp. 13648-13657 (2022)"
+ }
+ ]
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 131,
+ 457,
+ 481,
+ 479
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 457,
+ 481,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 457,
+ 481,
+ 479
+ ],
+ "type": "text",
+ "content": "49. Weyand, T., Kostrikov, I., Philbin, J.: Planet-photo geolocation with convolutional neural networks. In: ECCV. pp. 37-55 (2016)"
+ }
+ ]
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 131,
+ 479,
+ 481,
+ 501
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 479,
+ 481,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 479,
+ 481,
+ 501
+ ],
+ "type": "text",
+ "content": "50. Wilson, D., Zhang, X., Sultani, W., Wshah, S.: Image and object geo-localization. International Journal of Computer Vision pp. 1-43 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 131,
+ 502,
+ 481,
+ 522
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 502,
+ 481,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 502,
+ 481,
+ 522
+ ],
+ "type": "text",
+ "content": "51. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision-language models. In: CVPR. pp. 16816-16825 (2022)"
+ }
+ ]
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 131,
+ 523,
+ 481,
+ 544
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 523,
+ 481,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 523,
+ 481,
+ 544
+ ],
+ "type": "text",
+ "content": "52. Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337-2348 (2022)"
+ }
+ ]
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 131,
+ 545,
+ 481,
+ 577
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 545,
+ 481,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 545,
+ 481,
+ 577
+ ],
+ "type": "text",
+ "content": "53. Zhu, S., Yang, L., Chen, C., Shah, M., Shen, X., Wang, H.: R2former: Unified retrieval and reranking transformer for place recognition. In: CVPR. pp. 19370-19380 (2023)"
+ }
+ ]
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 131,
+ 578,
+ 481,
+ 599
+ ],
+ "type": "ref_text",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 131,
+ 578,
+ 481,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 131,
+ 578,
+ 481,
+ 599
+ ],
+ "type": "text",
+ "content": "54. Zhu, Y., Wang, J., Xie, L., Zheng, L.: Attention-based pyramid aggregation network for visual place recognition. In: ACM MM. pp. 99-107 (2018)"
+ }
+ ]
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "sub_type": "ref_text"
+ }
+ ],
+ "discarded_blocks": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "header",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 391,
+ 91,
+ 447,
+ 100
+ ],
+ "type": "text",
+ "content": "AddressCLIP"
+ }
+ ]
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 470,
+ 91,
+ 481,
+ 100
+ ],
+ "type": "page_number",
+ "angle": 0,
+ "lines": [
+ {
+ "bbox": [
+ 470,
+ 91,
+ 481,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 470,
+ 91,
+ 481,
+ 100
+ ],
+ "type": "text",
+ "content": "17"
+ }
+ ]
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "page_size": [
+ 612,
+ 792
+ ],
+ "page_idx": 16
+ }
+ ],
+ "_backend": "vlm",
+ "_version_name": "2.6.4"
+}
\ No newline at end of file
diff --git a/2024/AdvDiff_ Generating Unrestricted Adversarial Examples using Diffusion Models/e9674dff-c9d7-4719-bac8-0759677069f3_content_list.json b/2024/AdvDiff_ Generating Unrestricted Adversarial Examples using Diffusion Models/e9674dff-c9d7-4719-bac8-0759677069f3_content_list.json
new file mode 100644
index 0000000000000000000000000000000000000000..9372e84255f20538bc4ad3040bfc48bd1edcd56c
--- /dev/null
+++ b/2024/AdvDiff_ Generating Unrestricted Adversarial Examples using Diffusion Models/e9674dff-c9d7-4719-bac8-0759677069f3_content_list.json
@@ -0,0 +1,1795 @@
+[
+ {
+ "type": "text",
+ "text": "AdvDiff: Generating Unrestricted Adversarial Examples using Diffusion Models",
+ "text_level": 1,
+ "bbox": [
+ 241,
+ 140,
+ 763,
+ 186
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Xuelong Dai $^{1}$ , Kaisheng Liang $^{1}$ , and Bin Xiao $^{1}$",
+ "bbox": [
+ 313,
+ 210,
+ 687,
+ 227
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "The Hong Kong Polytechnic University \nxuelong.dai@connect.polyu.hk, cskliang@comp.polyu.edu.hk, \nb.xiao@polyu.edu.hk",
+ "bbox": [
+ 279,
+ 239,
+ 720,
+ 281
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Abstract. Unrestricted adversarial attacks present a serious threat to deep learning models and adversarial defense techniques. They pose severe security problems for deep learning applications because they can effectively bypass defense mechanisms. However, previous attack methods often directly inject Projected Gradient Descent (PGD) gradients into the sampling of generative models, which are not theoretically provable and thus generate unrealistic examples by incorporating adversarial objectives, especially for GAN-based methods on large-scale datasets like ImageNet. In this paper, we propose a new method, called AdvDiff, to generate unrestricted adversarial examples with diffusion models. We design two novel adversarial guidance techniques to conduct adversarial sampling in the reverse generation process of diffusion models. These two techniques are effective and stable in generating high-quality, realistic adversarial examples by integrating gradients of the target classifier interpretably. Experimental results on MNIST and ImageNet datasets demonstrate that AdvDiff is effective in generating unrestricted adversarial examples, which outperforms state-of-the-art unrestricted adversarial attack methods in terms of attack performance and generation quality.",
+ "bbox": [
+ 261,
+ 314,
+ 743,
+ 565
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "Keywords: Unrestricted Adversarial Attacks $\\cdot$ Diffusion Models $\\cdot$ Interpretable Adversarial Diffusion Sampling",
+ "bbox": [
+ 259,
+ 575,
+ 740,
+ 604
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "1 Introduction",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 628,
+ 375,
+ 645
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "While the deep learning (DL) community continues to explore the wide range of applications of DL models, researchers [37] have demonstrated that these models are highly susceptible to deception by adversarial examples. Adversarial examples are generated by adding perturbations to clean data. The perturbed examples can deceive DL classifiers with high confidence while remaining imperceptible to humans. Many strong attack methods [3,8,13,21,22,25] are proposed and investigated to improve the robustness of DL models.",
+ "bbox": [
+ 212,
+ 657,
+ 787,
+ 763
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "In contrast to existing perturbation-based adversarial attacks, Song et al. [35] found that using a well-trained generative adversarial network with an auxiliary classifier (AC-GAN) [29] can directly generate new adversarial examples without perturbing the clean data. These newly generated examples are considered unrestricted as they are obtained by optimizing input noise vectors",
+ "bbox": [
+ 212,
+ 763,
+ 787,
+ 839
+ ],
+ "page_idx": 0
+ },
+ {
+ "type": "text",
+ "text": "without any norm restrictions. Compared to traditional adversarial examples, unrestricted adversarial examples [9,30] are more aggressive against current adversarial defenses. A malicious adversary can also generate an unlimited number of unrestricted adversarial examples using a trained GAN.",
+ "bbox": [
+ 212,
+ 146,
+ 782,
+ 205
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Diffusion models [17] are likelihood-based generative models proposed recently, which emerged as a strong competitor to GANs. Diffusion models have outperformed GANs for image synthesis tasks [12, 19, 31]. Compared with GAN models, diffusion models are more stable during training and provide better distribution coverage. Diffusion models contain two processes: a forward diffusion process and a reverse generation process. The forward diffusion process gradually adds Gaussian noise to the data and eventually transforms it into noise. The reverse generation process aims to recover the data from the noise by a denoising-like technique. A well-trained diffusion model is capable of generating images with random noise input. Similar to GAN models, diffusion models can achieve adversarial attacks by incorporating adversarial objectives [4-6].",
+ "bbox": [
+ 212,
+ 210,
+ 784,
+ 377
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "GAN-based unrestricted adversarial attacks often exhibit poor performance on high-quality datasets, particularly in terms of visual quality, because they directly add the PGD perturbations to the GAN latents without theoretic supports. These attacks tend to generate low-quality adversarial examples compared to benign GAN examples [35]. Therefore, these attacks are not imperceptible among GAN synthetic data. Diffusion models, however, offer state-of-the-art generation performance [12] on challenging datasets like LSUN [41] and ImageNet [10]. The conditional diffusion models can generate images based on specific conditions by sampling from a perturbed conditional Gaussian noise, which can be carefully modified with adversarial objectives. These properties make diffusion models more suitable for conducting unrestricted adversarial attacks. Nevertheless, existing adversarial attack methods using diffusion models [4-6] adopt similar PGD perturbations to the sample in each reverse generation process, making them generate relatively low-quality adversarial examples.",
+ "bbox": [
+ 212,
+ 380,
+ 785,
+ 592
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "In this paper, we propose a novel and interpretable unrestricted adversarial attack method called AdvDiff that utilizes diffusion models for adversarial examples generation, as shown in Figure 1. Specifically, AdvDiff uses a trained conditional diffusion model to conduct adversarial attacks with two new adversarial guidance techniques. 1) During the reverse generation process, we gradually add adversarial guidance by increasing the likelihood of the target attack label. 2) We perform the reverse generation process multiple times, adding adversarial prior knowledge to the initial noise with the noise sampling guidance.",
+ "bbox": [
+ 212,
+ 595,
+ 785,
+ 715
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "text",
+ "text": "Our theoretical analysis indicates that these adversarial guidance techniques can effectively craft adversarial examples by the reverse generation process with adversarial conditional sampling. Furthermore, the sampling of AdvDiff benefits from stable and high sample quality of the diffusion models sampling, which leads to the generation of realistic unrestricted adversarial examples. Through extensive experiments conducted on two datasets, i.e., the high-quality dataset ImageNet, and the small, robust dataset MNIST, we have observed a significant improvement in the attack performance using AdvDiff with diffusion models.",
+ "bbox": [
+ 212,
+ 719,
+ 785,
+ 839
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "page_number",
+ "text": "2",
+ "bbox": [
+ 217,
+ 114,
+ 228,
+ 126
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "header",
+ "text": "X. Dai et al.",
+ "bbox": [
+ 271,
+ 114,
+ 357,
+ 127
+ ],
+ "page_idx": 1
+ },
+ {
+ "type": "image",
+ "img_path": "images/5da25d31598591bddc1684bd4891b2c06ccf5c64447a1bad730f49aff09ce165.jpg",
+ "image_caption": [
+ "Fig.1: The two new guidance techniques in our AdvDiff to generate unrestricted adversarial examples. During the reverse generation process, the adversarial guidance is added at timestep $x_{t}$ , which injects the adversarial objective $y_{a}$ into the diffusion process. The noise sampling guidance modifies the original noise by increasing the conditional likelihood of $y_{a}$ ."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 233,
+ 148,
+ 769,
+ 290
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "These results prove that our proposed AdvDiff is more effective than previous unrestricted adversarial attack methods in conducting unrestricted adversarial attacks to generate high-fidelity and diverse examples without decreasing the generation quality.",
+ "bbox": [
+ 214,
+ 417,
+ 782,
+ 477
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "Our contributions can be summarized as follows:",
+ "bbox": [
+ 238,
+ 478,
+ 589,
+ 491
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "list",
+ "sub_type": "text",
+ "list_items": [
+ "- We propose AdvDiff, the new form unrestricted adversarial attack method that utilizes the reverse generation process of diffusion models to generate realistic adversarial examples.",
+ "- We design two new effective adversarial guidance techniques to the sampling process that incorporate adversarial objectives to the diffusion model without re-training the model. Theoretical analysis reveals that AdvDiff can generate unrestricted adversarial examples while preserving the high-quality and stable sampling of the conditional diffusion models.",
+ "- We perform extensive experiments to demonstrate that AdvDiff achieves an overwhelmingly better performance than GAN models on unrestricted adversarial example generation."
+ ],
+ "bbox": [
+ 225,
+ 506,
+ 784,
+ 672
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2 Preliminaries",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 696,
+ 380,
+ 712
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "In this section, we introduce the diffusion model and the classifier guidance for constructing our adversarial diffusion model.",
+ "bbox": [
+ 214,
+ 729,
+ 782,
+ 758
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "2.1 Diffusion Model",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 782,
+ 393,
+ 797
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "The Denoising Diffusion Probabilistic Model (DDPM) [17] is utilized to learn the Gaussian transitions from $p(x_{t}) = \\mathcal{N}(x_{t};0,\\mathbf{I})$ to recover the data $x_0\\sim q(x_0)$",
+ "bbox": [
+ 214,
+ 809,
+ 782,
+ 840
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "header",
+ "text": "AdvDiff",
+ "bbox": [
+ 676,
+ 114,
+ 733,
+ 126
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "page_number",
+ "text": "3",
+ "bbox": [
+ 774,
+ 116,
+ 784,
+ 126
+ ],
+ "page_idx": 2
+ },
+ {
+ "type": "text",
+ "text": "with a Markov chain. We call this denoising-like process the reverse generation process. Following the pre-defined $T$ time steps, DDPM obtains a sequence of noisy data $\\{x_{T - 1},\\ldots ,x_1\\}$ and finally recovers the data $x_0$ . It is defined as:",
+ "bbox": [
+ 212,
+ 146,
+ 787,
+ 191
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) := \\mathcal {N} \\left(x _ {t - 1}: \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) \\tag {1}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 348,
+ 205,
+ 785,
+ 222
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Conversely, the forward diffusion process leverages a fixed Markov chain to iteratively add Gaussian noise to the sampled data $x_0 \\sim q(x_0)$ according to the scheduling function $\\beta_1, \\ldots, \\beta_N$ . Specifically, it is defined as:",
+ "bbox": [
+ 212,
+ 233,
+ 784,
+ 280
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) := \\mathcal {N} \\left(x _ {t}: \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right) \\tag {2}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 366,
+ 292,
+ 785,
+ 310
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Training of the DDPM requires accurate inference of the mean value $\\mu_{\\theta}$ with a deep learning model in the reverse generation process. The objective is to learn the variational lower-bound (VLB) on $\\log p_{\\theta}(x_0)$ . In order to complete the training objective, we train the model $\\epsilon_{\\theta}$ to predict the added Gaussian noise in the forward diffusion process. The standard mean-squared error loss is adopted:",
+ "bbox": [
+ 212,
+ 321,
+ 782,
+ 398
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\mathcal {L} _ {D D P M} := E _ {t \\sim [ 1, T ], \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I})} \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) \\| ^ {2} \\tag {3}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 349,
+ 412,
+ 785,
+ 430
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Song et al. [34] proposed DDIM, which provides an alternative noising process without restricting to a Markov chain. DDIM can achieve a much faster sampling than DDPM with the same training as DDPM. We perform experiments on both DDPM and DDIM to demonstrate the usability of our AdvDiff. DDPM is adopted to introduce our method for simplicity.",
+ "bbox": [
+ 212,
+ 436,
+ 787,
+ 513
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.2 Classifier-Guided Guidance",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 537,
+ 488,
+ 551
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Dhariwal et al. [12] achieved conditional diffusion sampling by adopting a trained classifier. The conditional information is injected into the diffusion model by modifying the mean value $\\mu_{\\theta}(x_t,t)$ of the samples according to the gradient of the prediction of the target class $y$ by the trained classifier. They adopted log probability to calculate the gradient, and the mean value is given by:",
+ "bbox": [
+ 212,
+ 564,
+ 787,
+ 642
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {\\mu} _ {\\theta} \\left(x _ {t}, t\\right) = \\mu_ {\\theta} \\left(x _ {t}, t\\right) + s \\cdot \\nabla_ {x _ {t}} \\log p _ {\\phi} (y | x _ {t}) \\tag {4}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 357,
+ 654,
+ 785,
+ 670
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "where $s$ is the guidance scale.",
+ "bbox": [
+ 214,
+ 681,
+ 433,
+ 698
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "2.3 Classifier-Free Guidance",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 722,
+ 464,
+ 736
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "Ho et al. [18] recently proposed a new conditional diffusion model using classifier-free guidance that injects class information without adopting an additional classifier. The classifier-free guidance utilizes a conditional diffusion model $p_{\\theta}(x|y)$ for image synthesis with given labels. For effective training, they jointly train the unconditional diffusion model $p_{\\theta}(x|\\emptyset)$ and the conditional diffusion model $p_{\\theta}(x|y)$ , where the unconditional diffusion model is simply replacing the label",
+ "bbox": [
+ 212,
+ 748,
+ 787,
+ 842
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "page_number",
+ "text": "4",
+ "bbox": [
+ 217,
+ 114,
+ 228,
+ 126
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "header",
+ "text": "X. Dai et al.",
+ "bbox": [
+ 271,
+ 114,
+ 356,
+ 127
+ ],
+ "page_idx": 3
+ },
+ {
+ "type": "text",
+ "text": "information with $\\emptyset$ . Sampling is performed by pushing the model towards the latent space of $p_{\\theta}(x|y)$ and away from $p_{\\theta}(x|\\emptyset)$ :",
+ "bbox": [
+ 212,
+ 146,
+ 782,
+ 176
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} (x _ {t} | y) = \\epsilon_ {\\theta} (x _ {t} | \\emptyset) + w \\cdot (\\epsilon_ {\\theta} (x _ {t} | y) - \\epsilon_ {\\theta} (x _ {t} | \\emptyset)) \\tag {5}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 343,
+ 189,
+ 785,
+ 204
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $w$ is the weight parameter for class guidance and $\\emptyset$ is the empty set.",
+ "bbox": [
+ 214,
+ 215,
+ 751,
+ 229
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The idea of classifier-free guidance is inspired by the gradient of an implicit classifier $p^i(y|x) \\propto p(x|y)/p(x)$ , the gradient of the classifier would be:",
+ "bbox": [
+ 212,
+ 229,
+ 782,
+ 260
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\begin{array}{l} \\nabla_ {x} \\log p ^ {i} (y | x) \\propto \\nabla_ {x} \\log p (x | y) - \\nabla_ {x} \\log p (x) \\\\ \\propto \\epsilon_ {\\theta} \\left(x _ {t} | y\\right) - \\epsilon_ {\\theta} \\left(x _ {t} | \\emptyset\\right) \\tag {6} \\\\ \\end{array}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 354,
+ 272,
+ 784,
+ 306
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "The classifier-free guidance has a good capability of generating high-quality conditional images, which is critical for performing adversarial attacks. The generation of these images does not rely on a classification model and thus can better fit the conditional distribution of the data.",
+ "bbox": [
+ 212,
+ 316,
+ 782,
+ 376
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3 Adversarial Diffusion Sampling",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 400,
+ 552,
+ 417
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "3.1 Rethinking Unrestricted Adversarial Examples",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 431,
+ 645,
+ 446
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Song et al. [35] presented a new form of adversarial examples called unrestricted adversarial examples (UAEs). These adversarial examples are not generated by adding perturbations over the clean data but are directly generated by any generative model. UAEs can be viewed as false negative errors in the classification tasks, and they can also bring server security problems to deep learning models. These generative-based UAEs can be formulated as:",
+ "bbox": [
+ 212,
+ 455,
+ 782,
+ 546
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nA _ {\\mathrm {U A E}} \\triangleq \\{x \\in \\mathcal {G} \\left(z _ {\\mathrm {a d v}}, y\\right) | y \\neq f (x) \\} \\tag {7}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 380,
+ 561,
+ 784,
+ 578
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "where $f(\\cdot)$ is the target model for unrestricted adversarial attacks. The unrestricted adversarial attacks aim to generate UAEs that fool the target model while still can be visually perceived as the image from ground truth label $y$ .",
+ "bbox": [
+ 212,
+ 583,
+ 782,
+ 628
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "Previous UAE works adopt GAN models for the generation of UAEs, and these works perturb the GAN latents by maximizing the cross-entropy loss of the target model, i.e., $\\max_{z_{\\mathrm{adv}}}\\mathcal{L}(f(\\mathcal{G}(z_{\\mathrm{adv}},y)),y)$ . Ideally, the generated UAEs should guarantee similar generation quality to the samples crafted by standard $z$ because successful adversarial examples should be imperceptible to humans. In other words, UAEs should not be identified among the samples with adversarial latents and standard latents.",
+ "bbox": [
+ 212,
+ 628,
+ 784,
+ 733
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "text",
+ "text": "However, due to GAN's poor interpretability, there's no theoretical support on $z_{adv}$ that can craft UAEs with normally trained GANs. The generator of GAN is not trained with $z_{adv} = z + \\nabla \\mathcal{L}$ but only $z \\sim \\mathcal{N}(0, \\mathbf{I})$ . Therefore, GAN-based UAEs encounter a significant decrease in generation quality because samples with $z_{adv}$ are not well-trained compared with samples with $z \\sim \\mathcal{N}(0, \\mathbf{I})$ . Moreover, the GAN latents are sampled from low dimensional latent spaces. Therefore, GANs are extremely sensitive to the latent $z$ [23, 33]. If we inject",
+ "bbox": [
+ 212,
+ 734,
+ 784,
+ 840
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "header",
+ "text": "AdvDiff",
+ "bbox": [
+ 676,
+ 114,
+ 730,
+ 126
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "page_number",
+ "text": "5",
+ "bbox": [
+ 774,
+ 116,
+ 784,
+ 126
+ ],
+ "page_idx": 4
+ },
+ {
+ "type": "image",
+ "img_path": "images/b014ae852a983804fe88f6a7063026e4aa47f5311d4a36bf77204d5ad2f883c6.jpg",
+ "image_caption": [
+ "Fig. 2: Unrestricted adversarial examples generated by the diffusion model. The generated adversarial examples should be visually indistinguishable from clean data with label $y$ but wrongly classified by the target classifier $f$ ."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 397,
+ 146,
+ 599,
+ 267
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "gradients of the classification results into GAN latents, GAN-based methods are more likely to generate flipped-label UAEs (images corresponding to the targeted attack label $y_{a}$ instead of the conditional generation label $y$ ) and distorted UAEs. However, these generation issues are hard to address only by attack success rate (ASR). In other words, even with a high ASR, some of the successful UAEs with GAN-based methods should be identified as failure cases for poor visual quality. However, such cases can not be reflected by ASR but can be evaluated by generation quality. All these problems may indicate that GAN models are not suitable for generative-based adversarial attacks.",
+ "bbox": [
+ 212,
+ 369,
+ 787,
+ 503
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "Diffusion models have shown better performance on image generation than GAN models [12]. They are log-likelihood models with interpretable generation processes. In this paper, we aim to generate UAEs by injecting the adversarial loss with theoretical proof and without sabotaging the benign generation process, where we increase the conditional likelihood on the target attack label by following the diffusion process. The perturbations are gradually injected with the backward generation process of the diffusion model by the same sample procedure. As shown in Figure 2, the diffusion model can sample images from the conditional distribution $p(x|y)$ . The samples from $p(x|y, f(x) \\neq y)$ are the adversarial examples that are misclassified by $f(\\cdot)$ . These examples also follow the data distribution $p(x|y)$ but on the other side of the label $y$ 's decision boundary of $f(\\cdot)$ . Moreover, the diffusion model's generation process takes multiple sampling steps. Thus, we don't need one strong perturbation to the latent like GAN-based methods. The AdvDiff perturbations at each step are unnoticeable, and perturbations are added to the high dimensional sampled data rather than low dimensional latents. Therefore, AdvDiff with diffusion models can preserve the generation quality and barely generates flipped-label or distorted UAEs.",
+ "bbox": [
+ 212,
+ 506,
+ 787,
+ 763
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "3.2 Adversarial Diffusion Sampling with Theoretical Support",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 784,
+ 732,
+ 800
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "There are several existing adversarial attack methods [4-6] that adopt diffusion models to generate adversarial examples. However, these methods still adopt",
+ "bbox": [
+ 214,
+ 809,
+ 785,
+ 840
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "page_number",
+ "text": "6",
+ "bbox": [
+ 217,
+ 114,
+ 228,
+ 126
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "header",
+ "text": "X. Dai et al.",
+ "bbox": [
+ 271,
+ 114,
+ 357,
+ 127
+ ],
+ "page_idx": 5
+ },
+ {
+ "type": "text",
+ "text": "PGD or I-FGSM gradients to perturb the diffusion process for constructing adversarial examples. As discussed earlier, the generation process of diffusion models is a specially designed sampling process from given distributions. Such adversarial gradients change the original generation process and can harm the generation quality of the diffusion model. Additionally, these methods fail to give a comprehensive discussion of the adversarial guidance with theoretical analysis. Therefore, we aim to design a general and interpretable method to generate adversarial examples using diffusion models without affecting the benign diffusion process.",
+ "bbox": [
+ 212,
+ 146,
+ 782,
+ 282
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "3.3 Adversarial Guidance",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 306,
+ 439,
+ 319
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Inspired by Dhariwal's work [12] that achieves the conditional image generation by classifier gradient guidance $\\nabla_{x_t}\\log p_\\phi (y|x_t)$ , we generate our UAEs with adversarial gradient guidance over the reverse generation process. Our attack aims at utilizing a conditional diffusion model $\\epsilon_{\\theta}(x_{t},y)$ to generate $x_0$ fits the ground truth label $y$ while deceiving the target classifier with $p_f(x_0)\\neq y$ . These generated samples are the false negative results in $p_f$ 's classification results.",
+ "bbox": [
+ 212,
+ 330,
+ 782,
+ 421
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Normally, we will obtain the images with label $y$ by following the standard reverse generation process with classifier-free guidance:",
+ "bbox": [
+ 212,
+ 422,
+ 782,
+ 452
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nx _ {t - 1} = \\mu \\left(x _ {t}, y\\right) + \\sigma_ {t} \\varepsilon \\tag {8}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 426,
+ 465,
+ 784,
+ 481
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "where $\\mu (x_{t},y)$ is the conditional mean value and $\\varepsilon$ is sampled from $\\varepsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ .",
+ "bbox": [
+ 212,
+ 492,
+ 782,
+ 508
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "Sampling by Equation 8, we obtain the samples with the generation process $p(x_{t-1}|x_t,y)$ . Following the above-mentioned definition of UAEs, we can get our adversarial examples by adding adversarial guidance to the standard reverse process, which is performing another sampling with the adversarial generation process $p(x_{t-1}^*|x_{t-1},f(x)\\neq y)$ . We find that specifying a target label for the adversarial generation process is more effective during experiments. Suggest the target label $y_a$ is the target for the adversarial attacks, the adversarial example is sampled by the following steps:",
+ "bbox": [
+ 212,
+ 508,
+ 784,
+ 628
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nx _ {t - 1} ^ {*} = x _ {t - 1} + \\sigma_ {t} ^ {2} s \\nabla_ {x _ {t - 1}} \\log p _ {f} \\left(y _ {a} \\mid x _ {t - 1}\\right) \\tag {9}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 361,
+ 642,
+ 784,
+ 657
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "where $s$ is the adversarial guidance scale. The derivation of Equation 9 is given in Appendix A. Intuitively, the adversarial guidance encourages the generation of samples with a higher likelihood of the target label.",
+ "bbox": [
+ 212,
+ 670,
+ 782,
+ 714
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "In practice, we utilize the classifier-free guidance to train a conditional diffusion model $\\epsilon_{\\theta}(\\cdot)$ as our basic generation model.",
+ "bbox": [
+ 212,
+ 715,
+ 782,
+ 744
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "3.4 Noise Sampling Guidance",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 768,
+ 473,
+ 782
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "text",
+ "text": "We can improve the reverse process by adding an adversarial label prior to the noise data $x_{T}$ . The UAEs are a subset of the dataset labeled with $y$ . They can be viewed as the conditional probability distribution with $p(x|y, f(x) = y_a)$ during",
+ "bbox": [
+ 212,
+ 794,
+ 782,
+ 840
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "header",
+ "text": "AdvDiff",
+ "bbox": [
+ 676,
+ 114,
+ 730,
+ 126
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "page_number",
+ "text": "7",
+ "bbox": [
+ 774,
+ 116,
+ 785,
+ 126
+ ],
+ "page_idx": 6
+ },
+ {
+ "type": "image",
+ "img_path": "images/1533fd941ba6f4c48111a4b9b265f10954bc49fb42a27344a3d9ea6200e5f042.jpg",
+ "image_caption": [
+ "Fig. 3: Adversarial examples on the MNIST dataset. Perturbation-based attack methods generate noise patterns to conduct attacks, while unrestricted adversarial attacks (U-GAN and AdvDiff) are imperceptible to the clean data."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 323,
+ 152,
+ 681,
+ 239
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "sampling, and $y_{a}$ is the target label for the adversarial attack. Therefore, we can add the adversarial label prior to $x_{t}$ with Bayes' theorem:",
+ "bbox": [
+ 214,
+ 343,
+ 782,
+ 375
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\n\\begin{array}{l} p \\left(x _ {T} \\mid y _ {a}\\right) = \\frac {p \\left(y _ {a} \\mid x _ {T}\\right) p \\left(x _ {T}\\right)}{p \\left(y _ {a}\\right)} = \\frac {p \\left(y _ {a} \\mid x _ {T} , x _ {0}\\right) p \\left(x _ {T} \\mid x _ {0}\\right)}{p \\left(y _ {a} \\mid x _ {0}\\right)} \\\\ = p \\left(x _ {T} \\mid x _ {0}\\right) e ^ {\\log p \\left(y _ {a} \\mid x _ {T}\\right) - \\log p \\left(y _ {a} \\mid x _ {0}\\right)} \\tag {10} \\\\ \\end{array}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 318,
+ 383,
+ 784,
+ 438
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "We can infer the $x_{t}$ with the adversarial prior by Equation 10, i.e.,",
+ "bbox": [
+ 240,
+ 450,
+ 720,
+ 465
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "equation",
+ "text": "\n$$\nx _ {T} = \\left(\\mu \\left(x _ {0}, y\\right) + \\sigma_ {t} \\varepsilon\\right) + \\bar {\\sigma} _ {T} ^ {2} a \\nabla_ {x _ {0}} \\log p _ {f} \\left(y _ {a} \\mid x _ {0}\\right) \\tag {11}\n$$\n",
+ "text_format": "latex",
+ "bbox": [
+ 336,
+ 477,
+ 784,
+ 494
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "where $a$ is the noise sampling guidance scale. See Appendix B for detailed proof.",
+ "bbox": [
+ 214,
+ 506,
+ 782,
+ 521
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Equation 11 is similar to Equation 9 as they both add adversarial guidance to the reverse generation process. However, the noise sampling guidance is added to $x_{T}$ according to the final classification gradient $\\nabla_{x_0}\\log p_f(y_a|x_0)$ , which provides a strong adversarial guidance signal directly to the initial input of the generative model. The gradient of Equation 11 is effective as it reflects the eventual classification result of the target classifier.",
+ "bbox": [
+ 212,
+ 522,
+ 782,
+ 611
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "3.5 Training-Free Adversarial Attack",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 635,
+ 532,
+ 650
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "The proposed adversarial attack does not require additional modification on the training of the diffusion model. The adversarial examples are sampled by using Algorithm 1 over the trained classifier-free diffusion model $\\epsilon_{\\theta}(\\cdot)$ . We give the AdvDiff algorithm on DDIM in the Appendix.",
+ "bbox": [
+ 214,
+ 662,
+ 782,
+ 723
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "4 Experiments",
+ "text_level": 1,
+ "bbox": [
+ 214,
+ 746,
+ 375,
+ 763
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "text",
+ "text": "Datasets and target models. We use two datasets for major evaluation: MNIST [11] and ImageNet [10]. MNIST is a 10-classes dataset consisting of handwritten numbers from 0 to 9. We adopt the MNIST dataset to evaluate our method for low-quality robust image generation. ImageNet is a large visual",
+ "bbox": [
+ 212,
+ 779,
+ 785,
+ 840
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "page_number",
+ "text": "8",
+ "bbox": [
+ 217,
+ 114,
+ 228,
+ 126
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "header",
+ "text": "X. Dai et al.",
+ "bbox": [
+ 271,
+ 114,
+ 357,
+ 126
+ ],
+ "page_idx": 7
+ },
+ {
+ "type": "code",
+ "sub_type": "algorithm",
+ "code_caption": [
+ "Algorithm 1 DDPM Adversarial Diffusion Sampling"
+ ],
+ "code_body": "Require: $y_{a}$ : target label for adversarial attack \nRequire: $y$ : ground truth class label \nRequire: $s,a$ : adversarial guidance scale \nRequire: $w$ : classification guidance scale \nRequire: $N$ : noise sampling guidance steps \nRequire: $T$ : reverse generation process timestep \n1: $x_{T}\\sim \\mathcal{N}(0,\\mathbf{I})$ \n2: $x_{adv} = \\emptyset$ \n3: for $i = 1\\dots N$ do \n4: for $t = T,\\ldots ,1$ do \n5: $\\tilde{\\epsilon}_t = (1 + w)\\epsilon_\\theta (x_t,y) - w\\epsilon_\\theta (x_t)$ \n6: Classifier-free sampling $x_{t - 1}$ with $\\tilde{\\epsilon}_t$ \n7: Input $x_{t - 1}$ to target model and get the gradient $\\log p_f(y_a|x_{t - 1})$ \n8: $x_{t - 1}^{*} = x_{t - 1} + \\sigma_{t}^{2}s\\nabla_{x_{t - 1}}\\log p_{f}(y_{a}|x_{t - 1})$ \n9: end for \n10: Obtain classification result from $f(x_0)$ \n11: Compute the gradient with $\\log p_f(y_a|x_0)$ \n12: Update $x_{T}$ by $x_{T} = x_{T} + \\bar{\\sigma}_{T}^{2}a\\nabla_{x_{0}}\\log p_{f}(y_{a}|x_{0})$ \n13: $x_{adv}\\gets x_0$ if $f(x_0) = y_a$ \n14: end for \n15: return $x_{adv}$",
+ "bbox": [
+ 215,
+ 164,
+ 718,
+ 458
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "database with 1000 object classes and is used for the high-quality generation task. For target classifiers, we adopt simple LeNet5 [20], and ResNet18 [15] for the MNIST dataset, and the widely-used ResNet50 [15] and WideResNet50-2 [42] for the ImageNet dataset.",
+ "bbox": [
+ 212,
+ 489,
+ 784,
+ 551
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "Comparisons. It is not applicable to give a clear comparison between perturbation attacks and unrestricted attacks because perturbation attacks have the corresponding ground truth while unrestricted attacks do not. We mainly compare our method with the unrestricted adversarial attack U-GAN [35] and give the discussion with the AutoAttack [8], PGD [25], BIM [13], and C&W [3] perturbation-based attacks under norm $\\ell_{\\mathrm{inf}} = 8 / 255$ . For U-GAN, We adopt AC-GAN [29] for the MNIST dataset, and SAGAN [43] and BigGAN [2] for the ImageNet dataset, as AC-GAN has shown poor performance on ImageNet. We use the official code from DiffAttack [4] and implement AdvDiffuser by ourselves [5] for comparisons. We do not compare with Chen et al. [6], because they use a similar method as DiffAttack and without official code. Because existing diffusion model attacks are all untargeted attacks, we include the untargeted version of AdvDiff for a clear comparison, which is represented by \"AdvDiff-Untargeted\".",
+ "bbox": [
+ 212,
+ 551,
+ 787,
+ 762
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "text",
+ "text": "Implementation details. Because our adversarial diffusion sampling does not require additional training to the original diffusion model, we use the pretrained diffusion model in our experiment. We adopt DDPM [17] with classifier-free guidance for the MNIST dataset and Latent Diffusion Model (LDM) [31] with DDIM sampler for the ImageNet dataset. For MNIST dataset, we use $N =$",
+ "bbox": [
+ 212,
+ 763,
+ 787,
+ 839
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "header",
+ "text": "AdvDiff",
+ "bbox": [
+ 676,
+ 114,
+ 733,
+ 127
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "page_number",
+ "text": "9",
+ "bbox": [
+ 774,
+ 114,
+ 785,
+ 126
+ ],
+ "page_idx": 8
+ },
+ {
+ "type": "image",
+ "img_path": "images/8ac12a995e014175a5b10aa7691ad5b539c234a58f2b2e94a00ac54fadad2f6d.jpg",
+ "image_caption": [
+ "Fig. 4: Comparisons of unrestricted adversarial attacks between GANs and diffusion models on two datasets.. Left: generated samples from U-GAN (BigGAN for ImageNet dataset). Right: generated samples from AdvDiff. We generate unrestricted adversarial examples on the MNIST \"0\" label and ImageNet \"mushroom\" label. U-GAN is more likely to generate adversarial examples with the target label, i.e., examples with red font. However, AdvDiff tends to generate the \"false negative\" samples by the target classifier by combing features from the target label."
+ ],
+ "image_footnote": [],
+ "bbox": [
+ 223,
+ 148,
+ 477,
+ 340
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "image",
+ "img_path": "images/d23d31f87fe4481def9d3667cc056677692868157ba4d6579cd6c9e81ea7f2c9.jpg",
+ "image_caption": [],
+ "image_footnote": [],
+ "bbox": [
+ 496,
+ 150,
+ 781,
+ 345
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "10, $s = 0.5$ , and $a = 1.0$ . And $N = 5$ , $s = 0.7$ , and $a = 0.5$ for ImageNet dataset. More details and experiments are given in the Appendix.",
+ "bbox": [
+ 212,
+ 501,
+ 784,
+ 532
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "Evaluation metrics. We utilize the top-1 classification result to evaluate the Attack Success Rate (ASR) on different attack methods under untargeted attack settings. As discussed earlier, GAN-based UAEs often encounter severe generation quality drops compared to benign GAN samples. Therefore, we give comparisons of generation performance on ImageNet to evaluate the attack performance of different UAEs in imperceptibly. The results are averaged with five runs. We use ResNet50 as the target model for default settings.",
+ "bbox": [
+ 212,
+ 532,
+ 787,
+ 638
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "4.1 Attack Performance",
+ "text_level": 1,
+ "bbox": [
+ 215,
+ 661,
+ 428,
+ 676
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "text",
+ "text": "MNIST We show the attack success rate against the normally trained model and adversarially trained model [25] in the MNIST dataset. All the selected adversarial attacks achieve over $90\\%$ attack success rate against the normally trained model. The adversarially trained model can effectively defend against perturbation-based adversarial attacks for their noise-like perturbation generation patterns, as reported in Table 1. However, the UAEs obviously perform better with their non-perturbed image generation. Despite the fact that the unrestricted attack can break through the adversarial defenses, the crafted adversarial examples should also be imperceptible to humans for a reasonably successful attack. The visualized adversarial examples in Figure 3 show that the",
+ "bbox": [
+ 212,
+ 688,
+ 787,
+ 840
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "page_number",
+ "text": "10",
+ "bbox": [
+ 217,
+ 114,
+ 235,
+ 126
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "header",
+ "text": "X. Dai et al.",
+ "bbox": [
+ 271,
+ 114,
+ 354,
+ 126
+ ],
+ "page_idx": 9
+ },
+ {
+ "type": "table",
+ "img_path": "images/b974354147bad58897878a73fd6915e3126d4d3856be56224547835a58f60a8a.jpg",
+ "table_caption": [
+ "Table 1: The attack success rate on MNIST dataset."
+ ],
+ "table_footnote": [],
+ "table_body": "