{"notes":[{"id":"retail-product/checksort","category":"retail-product","legacyPath":"retail-product/2306-checksort","legacySlug":"2306-checksort","legacyPaths":[],"title":"CheckSORT","subtitle":"Steadfast Trajectories","author":"Z. Yuan","paperDate":"2023","publishedOn":"2025-07-21","paperTitle":"CheckSORT: Refined Synthetic Data Combination and Optimized SORT for Automatic Retail Checkout","paperUrl":"https://openaccess.thecvf.com/content/CVPR2023W/AICity/html/Shi_CheckSORT_Refined_Synthetic_Data_Combination_and_Optimized_SORT_for_Automatic_CVPRW_2023_paper.html","bibliography":{"title":"CheckSORT: Refined Synthetic Data Combination and Optimized SORT for Automatic Retail Checkout","authors":["Ziqiang Shi","Zhongling Liu","Liu Liu","Rujie Liu","Takuma Yamamoto","Xiaoyu Mi","Daisuke Uchida"],"url":"https://openaccess.thecvf.com/content/CVPR2023W/AICity/html/Shi_CheckSORT_Refined_Synthetic_Data_Combination_and_Optimized_SORT_for_Automatic_CVPRW_2023_paper.html","pdfUrl":"https://openaccess.thecvf.com/content/CVPR2023W/AICity/papers/Shi_CheckSORT_Refined_Synthetic_Data_Combination_and_Optimized_SORT_for_Automatic_CVPRW_2023_paper.pdf","codeUrl":null,"preprint":null,"publication":{"venue":"CVPR Workshops","year":2023,"kind":"workshop","url":"https://openaccess.thecvf.com/content/CVPR2023W/AICity/html/Shi_CheckSORT_Refined_Synthetic_Data_Combination_and_Optimized_SORT_for_Automatic_CVPRW_2023_paper.html"},"citation":"@inproceedings{retailproductchecksort2023,\n  title = {{CheckSORT: Refined Synthetic Data Combination and Optimized SORT for Automatic Retail Checkout}},\n  author = {Ziqiang Shi and Zhongling Liu and Liu Liu and Rujie Liu and Takuma Yamamoto and Xiaoyu Mi and Daisuke Uchida},\n  year = {2023},\n  booktitle = {CVPR Workshops},\n  url = {https://openaccess.thecvf.com/content/CVPR2023W/AICity/html/Shi_CheckSORT_Refined_Synthetic_Data_Combination_and_Optimized_SORT_for_Automatic_CVPRW_2023_paper.html}\n}\n","citationPath":"/content/papers/citations/retail-product/checksort.8a6053659705.bib"},"thumbnail":{"src":"/content/papers/assets/1fa487277c1bfd2ba8580282905f4cc6b2e859bc5ab33aeb7554bffadd4b0aca.jpg","width":1004,"height":660},"locale":"en","path":"/content/papers/en/retail-product/checksort.266770ea6d9c.md","headings":[{"id":"problem-definition","text":"Problem Definition","depth":2},{"id":"solution","text":"Solution","depth":2},{"id":"synthetic-data-optimization","text":"Synthetic Data Optimization","depth":3},{"id":"model-architecture","text":"Model Architecture","depth":3},{"id":"sort-algorithm-overview","text":"SORT Algorithm Overview","depth":3},{"id":"decomposed-kalman-filter","text":"Decomposed Kalman Filter","depth":3},{"id":"dynamic-feature-sequences","text":"Dynamic Feature Sequences","depth":3},{"id":"postprocessing-optimization","text":"Postprocessing Optimization","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"impact-of-synthetic-data","text":"Impact of Synthetic Data","depth":3},{"id":"synthetic-vs-real-backgrounds","text":"Synthetic vs. Real Backgrounds","depth":3},{"id":"checksort-performance-analysis","text":"CheckSORT Performance Analysis","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"multi-object-tracking/oc-sort","relation":"comparison","topic":"Tracking objects and checkout items"}]},{"id":"multi-object-tracking/oc-sort","category":"multi-object-tracking","legacyPath":"multi-object-tracking/2203-oc-sort","legacySlug":"2203-oc-sort","legacyPaths":[],"title":"OC-SORT","subtitle":"Breaking Free from the Cage of Kalman","author":"Z. Yuan","paperDate":"2022-03-27","publishedOn":"2025-10-04","paperTitle":"Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking","paperUrl":"https://arxiv.org/abs/2203.14360","bibliography":{"title":"Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking","authors":["Jinkun Cao","Jiangmiao Pang","Xinshuo Weng","Rawal Khirodkar","Kris Kitani"],"url":"https://arxiv.org/abs/2203.14360","pdfUrl":"https://arxiv.org/pdf/2203.14360","codeUrl":null,"preprint":{"id":"2203.14360","submittedOn":"2022-03-27","url":"https://arxiv.org/abs/2203.14360"},"publication":{"venue":"CVPR","year":2023,"kind":"conference","url":"https://arxiv.org/abs/2203.14360"},"citation":"@misc{multiobjecttrackingocsort2022,\n  title = {{Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking}},\n  author = {Jinkun Cao and Jiangmiao Pang and Xinshuo Weng and Rawal Khirodkar and Kris Kitani},\n  year = {2022},\n  eprint = {2203.14360},\n  archivePrefix = {arXiv},\n  url = {https://arxiv.org/abs/2203.14360}\n}\n","citationPath":"/content/papers/citations/multi-object-tracking/oc-sort.c4de3a8ba99e.bib"},"thumbnail":{"src":"/content/papers/assets/d6917fdb4727ca6a2fe2bdd05b5cba09c597767f7ef2c6762680fd861ed95548.jpg","width":1224,"height":536},"locale":"en","path":"/content/papers/en/multi-object-tracking/oc-sort.4359a1baa38a.md","headings":[{"id":"defining-the-problem","text":"Defining the problem","depth":2},{"id":"rethinking-sorts-limitations","text":"Rethinking SORT's limitations","depth":3},{"id":"why-a-track-gets-lost-again-right-after-being-recovered","text":"Why a track gets lost again right after being recovered","depth":3},{"id":"solving-the-problem","text":"Solving the problem","depth":2},{"id":"observation-centric-re-update-oru","text":"Observation-Centric Re-Update (ORU)","depth":3},{"id":"observation-centric-momentum-ocm","text":"Observation-Centric Momentum (OCM)","depth":3},{"id":"observation-centric-recovery-ocr","text":"Observation-Centric Recovery (OCR)","depth":3},{"id":"discussion","text":"Discussion","depth":2},{"id":"dancetrack-results","text":"DanceTrack results","depth":3},{"id":"mot17-results","text":"MOT17 results","depth":3},{"id":"mot20-results","text":"MOT20 results","depth":3},{"id":"kitti-results","text":"KITTI results","depth":3},{"id":"module-contribution-analysis","text":"Module contribution analysis","depth":3},{"id":"how-oru-generates-the-virtual-trajectory","text":"How ORU generates the virtual trajectory","depth":3},{"id":"choosing-δt-for-ocm","text":"Choosing Δt for OCM","depth":3},{"id":"speed","text":"Speed","depth":3},{"id":"afterward","text":"Afterward","depth":3},{"id":"conclusion","text":"Conclusion","depth":2}],"anchorAliases":{},"guide":null,"relatedReading":[{"id":"retail-product/checksort","relation":"comparison","topic":"Tracking objects and checkout items"}]}],"text":"Remember the DeepACO system we saw from the 2022 AI CITY CHALLENGE Track 4?\n\nNow let's look at the champion paper of the 2023 AI CITY CHALLENGE.\n\n## Problem Definition\n\nSelf-checkout is gradually becoming a standard in retail environments. Whether in unmanned stores or the digital transformation of traditional markets, the goal is to use image recognition systems to **quickly and accurately identify products and quantities** without interrupting the user’s flow.\n\nThis task is generally formalized in computer vision as a multi-module system combining \"multi-class product detection, classification, and counting.\"\n\nAt first glance, this task seems like simply combining existing object detection, tracking, and classification techniques.\n\nHowever, when actually implementing it, many issues arise.\n\nThis paper organizes its image-based automatic checkout system into five modules:\n\n\n\n![pipeline](/content/papers/assets/1fa487277c1bfd2ba8580282905f4cc6b2e859bc5ab33aeb7554bffadd4b0aca.jpg)\n\n\n\n1. **Preprocessing**: How to generate training samples closer to real shots from clean synthetic data?\n2. **Object Detection**: How to locate products and hands, and obtain masks for further processing?\n3. **Object Tracking**: How to track short-lived and possibly overlapping product trajectories?\n4. **Product Classification**: How to improve recognition accuracy and adapt to changes in viewpoint, occlusion, and lighting?\n5. **Postprocessing**: How to integrate temporary results across multiple frames into a final checkout product list?\n\nThe authors believe each part must be carefully designed to meet practical deployment needs.\n\n## Solution\n\nIn automatic checkout tasks, perhaps the most difficult part is obtaining representative training data, which not only involves on-site shooting but also requires labor-intensive frame-by-frame annotation.\n\nThus, the authors start with **synthetic data optimization**, proposing a controllable data synthesis pipeline that generates a large variety of synthetic images.\n\n### Synthetic Data Optimization\n\nSpecifically, participants receive clean synthetic images and masks for each product, then generate backgrounds and composite them to create training synthetic images.\n\nThis process may seem simple, but the competition results show why it deserves careful attention.\n\nHow products and backgrounds are composited strongly affects the training distribution and resulting performance. To enhance realism and diversity of synthetic images, the authors designed a controllable data synthesis pipeline centered around three hyperparameters:\n\n- **Number of products**\n- **Occlusion level (IoU)**\n- **Scaling size**\n\nThese three are interdependent: too many products increase occlusion, while too small scaling may cause distortion.\n\nTherefore, users can set parameter ranges, for example:\n\n- Product count: 1 to 6\n- Occlusion upper limit: IoU ≤ 0.3\n- Scaling ratio: 0.15 to 0.5\n\nThrough systematic combinations, a large variety of synthetic data is produced.\n\nAblations compare occlusion limits and scale ranges; the best values depend on the tested configuration, as discussed below.\n\nBesides parameter tuning, the authors compared impacts of different background sources.\n\nAlthough external data use was restricted by the official rules, the authors implemented a version with **real background images composited with products** within allowed limits. The real backgrounds yielded higher F1 in the configurations tested in this paper.\n\nA useful interpretation of these experiments is:\n\n> **In these checkout experiments, real photographed backgrounds helped more than the synthetic backgrounds used for comparison.**\n\nRegarding data usage strategy, the synthetic data was applied to train two major models:\n\n1. **Product Detection**: Fine-tuning the COCO pretrained **DetectoRS** model, which supports both detection and segmentation, to locate products and provide hand masks.\n2. **Product Classification**: Extracting product images with varied backgrounds from synthetic data to fine-tune classification models.\n\n### Model Architecture\n\nCheckSORT’s architecture extends from DeepSORT and StrongSORT. Relevant papers can be referenced here:\n\n- [**[17.03] Simple Online and Realtime Tracking with a Deep Association Metric**](https://arxiv.org/abs/1703.07402)\n- [**[22.06] StrongSORT: Make DeepSORT Great Again**](https://arxiv.org/abs/2202.13514)\n\nThe authors introduced two key improvements into the tracking logic:\n\n1. **Decomposed Kalman Filter (DKF)**\n2. **Dynamic EMA Tracklet Gallery**\n\n### SORT Algorithm Overview\n\nTraditional SORT / DeepSORT / StrongSORT employ a two-stage \"Tracking-by-detection\" approach, where the input is detection results for each frame, and the output is corresponding product trajectories.\n\nInside the tracker, detections initialize tracklets, whose future positions are predicted via a **Kalman filter**. Tracklets maintain three states based on matching:\n\n- `tentative`: newly created tracks\n- `confirmed`: upgraded after consecutive correct matches\n- `deleted`: removed after long no-match periods\n\nPer frame, the processing steps are:\n\n1. **Predict**: From each tracklet’s history, predict its bounding box position and uncertainty in the current frame.\n2. **Match**: First match confirmed tracklets by appearance cosine distance and the Hungarian algorithm. Then use Mahalanobis distance between Kalman predictions and current detections for unconfirmed and still-unmatched tracklets.\n3. **Update**: Update the tracklet states using the results of both matching stages.\n\n### Decomposed Kalman Filter\n\nDeepSORT and the related baseline represent bounding boxes with this state; it is not the only possible Kalman-filter state:\n\n$$\n\\mathbf{x} = [x, y, a, h, \\dot{x}, \\dot{y}, \\dot{a}, \\dot{h}]\n$$\n\nwhere:\n\n- $(x, y)$: center coordinates\n- $a$: aspect ratio\n- $h$: height\n- \"$\\dot{\\ }$\": derivative (velocity)\n\nThe problem is that **motion patterns of center movement $(x, y)$ and appearance deformation $(a, h)$ differ significantly**, as shown below:\n\n\n\n![kalman filter](/content/papers/assets/1e01592d87f14b275be5c93920f025f5ef0a12940dc3326692c57986143c7ee6.jpg)\n\n\n\n- **Orange curve**: the comparatively smooth distance from the product center to the origin over time;\n- **Red curve**: often violent jumps in $(a, h)$ during product flipping.\n\nTherefore, the authors propose the Decomposed Kalman Filter (DKF), splitting the overall state into two independent Kalman computations:\n\n- **Position vector**:\n\n  $$\n  \\mathbf{p} = [x, y, \\dot{x}, \\dot{y}]\n  $$\n\n- **Shape vector**:\n\n  $$\n  \\mathbf{b} = [a, h, \\dot{a}, \\dot{h}]\n  $$\n\nBoth share the following transition matrix $A$ and observation matrix $H$:\n\n$$\nA =\n\\begin{bmatrix}\n1 & 0 & 1 & 0 \\\\\n0 & 1 & 0 & 1 \\\\\n0 & 0 & 1 & 0 \\\\\n0 & 0 & 0 & 1\n\\end{bmatrix}, \\quad\nH =\n\\begin{bmatrix}\n1 & 0 & 0 & 0 \\\\\n0 & 1 & 0 & 0\n\\end{bmatrix}\n$$\n\nFor matching, compare the $j$th detection’s position observation $\\mathbf z^{(p)}_j=(x_j,y_j)$ and shape observation $\\mathbf z^{(b)}_j=(a_j,h_j)$ with the $i$th tracklet’s predictions projected into two-dimensional observation space. $P_i,B_i$ are their respective projected covariances; the observation vectors do not include velocities.\n\n- **Position distance**:\n\n  $$\n  d^{(p)}(i,j) = (\\mathbf{z}^{(p)}_j - \\hat{\\mathbf{z}}^{(p)}_i)^\\top P_i^{-1} (\\mathbf{z}^{(p)}_j - \\hat{\\mathbf{z}}^{(p)}_i)\n  $$\n\n- **Shape distance**:\n\n  $$\n  d^{(b)}(i,j) = (\\mathbf{z}^{(b)}_j - \\hat{\\mathbf{z}}^{(b)}_i)^\\top B_i^{-1} (\\mathbf{z}^{(b)}_j - \\hat{\\mathbf{z}}^{(b)}_i)\n  $$\n\n### Dynamic Feature Sequences\n\nSince DeepSORT, trajectory matching has incorporated appearance features instead of relying solely on position.\n\nCheckSORT introduces an improved strategy:\n\n- **StrongSORT** uses EMA to smooth appearance features;\n- **CheckSORT** adopts a **dynamic, non-continuous feature gallery**, adding new features only when sufficiently different from historical ones to avoid redundancy.\n\nDefined as follows:\n\n- Each track $i$ has a gallery of normalized, EMA-smoothed appearance features:\n\n  $$\n  E_i = \\{ e_i^t \\}, \\quad t = 1 \\ldots L_i\n  $$\n\n- A new detection’s EMA feature is added only if it differs enough from those already in $E_i$;\n\n- Appearance distance is defined as:\n\n  $$\n  d^{(a)}(i,j) = \\min_{e\\in E_i}(1-e^\\top e_j)\n  $$\n\nHere $i$ indexes the track and $j$ the current detection, whose normalized feature is $e_j$. After distance-based gating, combine the position, shape, and appearance costs:\n\n$$\nc(i,j) = \\lambda_p \\cdot d^{(p)}(i,j) + \\lambda_b \\cdot d^{(b)}(i,j) + \\lambda_a \\cdot d^{(a)}(i,j)\n$$\n\nThis enables the Hungarian algorithm to achieve optimal matching by jointly considering position, shape, and appearance.\n\n### Postprocessing Optimization\n\nEven with high-quality matching results, CheckSORT performs a final \"trajectory correction\" step addressing three common error types:\n\n1. Short trajectories or those belonging to background classes → **deleted**\n2. Trajectory breaks longer than 0.5 seconds → **split** into two tracks\n3. Multiple tracks of the same class with time difference less than 3 seconds → **merged**\n\nThis part is heuristic and not a core contribution of the paper.\n\nThe figure below shows recognition differences before and after postprocessing:\n\n\n\n![post-process](/content/papers/assets/d29b4308ece50a46b2632aebebdc2c413fef4790380c3670885193654a561596.jpg)\n\n\n\n## Discussion\n\nThe authors observed that the **AI CITY CHALLENGE 2023** Track 4 task differs slightly from the previous year: in 2023, the task requires predicting the frame where each product **first appears**, whereas in 2022, the appearance timestamp was expressed in seconds.\n\nAlthough the authors intended to verify the algorithm directly on the 2023 test set, lack of ground truth forced them to manually annotate product appearance times (in seconds) on _TestA 2022_ and _TestA 2023_ as an alternative benchmark.\n\nThe evaluation follows 2022 rules, using **F1-score** as metric:\n\n- **TP (True Positive)**: correctly recognized objects in the region of interest;\n- **FN (False Negative)**: annotated objects not correctly recognized.\n\nAdditionally, the competition provided **116,500 synthetic images** and **116 3D product models**.\n\n### Impact of Synthetic Data\n\n\n\n![Performance comparison on TestA 2022 under different IoUs](/content/papers/assets/316ad1dbfdc035f765b3735b98ea812bfc297cf0fcb37576099ce2898e182e9e.jpg)\n\n\nCheckSORT performance on _TestA 2022_ under varying IoUs\n\n\n\n\n\n\n\n![Performance comparison on TestA 2023 under different IoUs](/content/papers/assets/cbc3195605298ada1327d786c7068695c75a4758691f4ce6a337306013854035.jpg)\n\n\nCheckSORT performance on _TestA 2023_ under varying IoUs\n\n\n\n\n\nThe authors designed controlled experiments to study two hyperparameters affecting performance: the **IoU upper bound** and the **scaling range**. With fixed scaling of 0.15–0.5, they observed recognition results under different IoU limits as shown above.\n\nAmong these discrete settings, an IoU cap of 0.25 gives the best F1 on both test sets (58.6% and 51.0%). The trend is not monotonic: 0.20 is worse than 0.10. The result supports including suitable occlusion rather than maximizing occlusion without limit. The tracker is fixed to DeepSORT here.\n\nNext, with IoU fixed at 0.1, the authors examined effects of varying scaling ranges:\n\n\n\n![Performance comparison on TestA 2022 under different scalings](/content/papers/assets/64b5dd36a4764809185323b26cc893cb054d27d8fbf998ba9140703ce75d2e66.jpg)\n\n\nCheckSORT performance on _TestA 2022_ under varying scaling ranges\n\n\n\n\n\n\n\n![Performance comparison on TestA 2023 under different scalings](/content/papers/assets/a2431127eb45fa65be8a2f0584769f595ce86c12bd82dc75b26c49bb119e6ef7.jpg)\n\n\nCheckSORT performance on _TestA 2023_ under varying scaling ranges\n\n\n\n\n\nAmong the three scale ranges compared, **0.15–0.55** gives the highest F1 on both test sets (58.2% and 50.2%). This is a sampling range, not a fixed scale factor of 0.55; the tracker is again fixed to DeepSORT.\n\n- Too large scaling causes insufficient samples;\n- Too small scaling mismatches real application scenarios, harming generalization.\n\n### Synthetic vs. Real Backgrounds\n\nBased on the best parameter combinations, the authors further evaluated the influence of background style.\n\nThey tested two settings — \"synthetic backgrounds\" and \"real backgrounds\" — on _TestA 2022_ and _2023_ datasets with results as follows:\n\n\n\n![Performance comparison under IoU≤0.5, scaling=0.4∼0.8.](/content/papers/assets/2f3464273dd17b48a0c203d1420b8b1dee01150be9fa95ebc22ac566bb95c6af.jpg)\n\n\nPerformance under IoU ≤ 0.5, scaling range 0.4–0.8\n\n\n\n\n\n\n\n![Performance comparison under IoU≤0.5, scaling=0.3∼0.75](/content/papers/assets/3a7f355f40ea5128b5fc6c58cc1ecaa44fa3624e94141aef126266d4834ebdb2.jpg)\n\n\nPerformance under IoU ≤ 0.5, scaling range 0.3–0.75\n\n\n\n\n\nReal backgrounds improve F1 in both configurations and test sets shown; individual precision or recall values need not all increase. Under IoU ≤ 0.5 and scaling 0.4–0.8, _TestA 2022_ F1 rises from 60.7% to 69.0%, an **8.3-percentage-point** gain; _TestA 2023_ rises from 53.7% to 54.9%. The tracker is fixed to DeepSORT here.\n\n### CheckSORT Performance Analysis\n\n\n\n![Performance comparison under different trackers with IoU≤0.5, scaling = 0.3∼0.75.](/content/papers/assets/4f1798cdc78888215029c392685158616aec8d533001512db4aeeea9f80d697d.jpg)\n\n\nComparison of trackers under IoU ≤ 0.5, scaling 0.3–0.75\n\n\n\n\n\n\n\n![Performance comparison under different trackers with IoU≤0.5, scaling = 0.4∼0.8.](/content/papers/assets/56e4818834eda2f80413aa72c12166f5c67bfa9fa0b4eeaa5ea6ffb21d319fd0.jpg)\n\n\nComparison of trackers under IoU ≤ 0.5, scaling 0.4–0.8\n\n\n\n\n\nFinally, the authors compared three trackers — **DeepSORT**, **StrongSORT**, and their proposed **CheckSORT** — across different scenarios.\n\nFor the two synthesis configurations and two test sets in these tables, **CheckSORT has higher F1 than DeepSORT and StrongSORT**. At scaling 0.3–0.75, _TestA 2022_ reports 70.3% / 69.0% / 67.8%, and _TestA 2023_ reports 62.1% / 58.2% / 60.0% (CheckSORT / DeepSORT / StrongSORT).\n\nAt paper submission, the official TestA leaderboard placed the authors’ team **5th**. The paper says the final ranking still awaited code evaluation on TestB; this is distinct from the manually annotated, seconds-based tests above.\n\n\n\n![AI City Challenge 2023 Track 4 Leaderboard](/content/papers/assets/36db316dbca75e3aca15982421e674cb0e4c06136d49344db9cf0714a4ff95bd.jpg)\n\n\n\n## Conclusion\n\nThe final results are well known.\n\n**CheckSORT’s team won the 2023 AI CITY CHALLENGE Track 4 final Test B evaluation with F1 = 0.8443.** This is distinct from the earlier Test A position above.\n\nCheckSORT devotes substantial effort to differences between training and test conditions: how synthetic products are placed, how much they overlap, and how to track them as they rotate. These choices directly affect performance in the checkout setting.\n\nThis paper addresses this core issue with two main solutions:\n\n- First, the authors designed a **synthetic data optimization pipeline**, controlling IoU upper bounds and scaling parameters to generate more representative training samples;\n- Second, they proposed a **decomposed filtering approach** at the filtering level to separate position and shape information, enabling stable tracking under rotation, flipping, and occlusion.\n\nIn the tested checkout scenes, these changes improve recognition and offer a useful basis for building a practical automatic checkout system.\n\n\n> [!NOTE]\n> \n> Interested readers can check out the leaderboard at the following link:\n> \n> - [**AI City Challenge 2023 Track 4 Leaderboard**](https://www.aicitychallenge.org/2023-challenge-winners/)\n> \n>   ![AI City Challenge 2023 Track 4 Leaderboard](/content/papers/assets/0f6e12754c47d35d3f1a014bc8b67725b29b70550e13e605c136da2a05ffc503.jpg)\n","imageSizes":{"/content/papers/assets/1fa487277c1bfd2ba8580282905f4cc6b2e859bc5ab33aeb7554bffadd4b0aca.jpg":{"width":1004,"height":660},"/content/papers/assets/1e01592d87f14b275be5c93920f025f5ef0a12940dc3326692c57986143c7ee6.jpg":{"width":1224,"height":596},"/content/papers/assets/d29b4308ece50a46b2632aebebdc2c413fef4790380c3670885193654a561596.jpg":{"width":704,"height":1080},"/content/papers/assets/316ad1dbfdc035f765b3735b98ea812bfc297cf0fcb37576099ce2898e182e9e.jpg":{"width":760,"height":212},"/content/papers/assets/cbc3195605298ada1327d786c7068695c75a4758691f4ce6a337306013854035.jpg":{"width":758,"height":218},"/content/papers/assets/64b5dd36a4764809185323b26cc893cb054d27d8fbf998ba9140703ce75d2e66.jpg":{"width":772,"height":216},"/content/papers/assets/a2431127eb45fa65be8a2f0584769f595ce86c12bd82dc75b26c49bb119e6ef7.jpg":{"width":780,"height":212},"/content/papers/assets/2f3464273dd17b48a0c203d1420b8b1dee01150be9fa95ebc22ac566bb95c6af.jpg":{"width":760,"height":336},"/content/papers/assets/3a7f355f40ea5128b5fc6c58cc1ecaa44fa3624e94141aef126266d4834ebdb2.jpg":{"width":748,"height":332},"/content/papers/assets/4f1798cdc78888215029c392685158616aec8d533001512db4aeeea9f80d697d.jpg":{"width":770,"height":254},"/content/papers/assets/56e4818834eda2f80413aa72c12166f5c67bfa9fa0b4eeaa5ea6ffb21d319fd0.jpg":{"width":784,"height":252},"/content/papers/assets/36db316dbca75e3aca15982421e674cb0e4c06136d49344db9cf0714a4ff95bd.jpg":{"width":792,"height":496},"/content/papers/assets/0f6e12754c47d35d3f1a014bc8b67725b29b70550e13e605c136da2a05ffc503.jpg":{"width":1224,"height":348}}}
