{"schemaVersion":2,"release":"wedraw-v4","totalClasses":413,"totalReferences":229745,"addedClasses":54,"dimensions":96,"modelSha256":"07c21e0105936a09cdc2ab1adb25b95badf242dbf4b3bb21ef56c37bac993c87","files":{"catalog.json":{"sha256":"22273005c85d4b8db3fd9cdc13bb65242a67cb6171d0c0e0beeda6a30110e17a","bytes":544630},"encoder.json":{"sha256":"07c21e0105936a09cdc2ab1adb25b95badf242dbf4b3bb21ef56c37bac993c87","bytes":764400},"prototypes.json":{"sha256":"c84e8ac616111f24484816f0831f258cba2a996080b68a757934c1713c951126","bytes":245749},"benchmark.json":{"sha256":"de25d80352ddabb1716b77fb3e1b0ea75cfa07ab9561f84011e624569d1c5bbd","bytes":3956},"matched-study.json":{"sha256":"6ecd696c169d77ee32ad522382f0a073d14e3dedb48519ba9b158faf3c5073c4","bytes":4567},"scoring-metric.mjs":{"sha256":"ae4d2914b3b2b2e70ed045cae79b89a27f45fde4a90008145e9854ae2b4d50e7","bytes":3632},"scoring-raster.mjs":{"sha256":"d1f65548352b8d53ba11c0a11aecce9274058267d9c44dd08ff2b0d179539ce5","bytes":1693},"scoring-features.mjs":{"sha256":"b9763ad0f5129c5969686b3391bd9dce824823a6bdce6b71d1c44a85c6509dec","bytes":2765}},"sources":{"common":{"name":"CommonSketch","creator":"Jiho Park, Sieun Choi, Jaeyoon Seo, Minho Sohn, Yeana Kim, Jihie Kim","url":"https://huggingface.co/datasets/ziiio/CommonSketch","kind":"human_freehand","license":"CC BY 4.0","licenseUrl":"https://creativecommons.org/licenses/by/4.0/"},"quickdraw":{"name":"The Quick, Draw! Dataset","creator":"Google, Inc. and Quick, Draw! players","url":"https://github.com/googlecreativelab/quickdraw-dataset","kind":"human_freehand","license":"CC BY 4.0","licenseUrl":"https://creativecommons.org/licenses/by/4.0/"},"seva":{"name":"SEVA: Human sketch production","creator":"Kushin Mukherjee, Holly Huey, Xuanchen Lu, Yael Vinker, Rio Aguina-Kang, Ariel Shamir, Judith E. Fan; Cognitive Tools Lab","url":"https://github.com/cogtoolslab/visual_abstractions_benchmarking_public2023","license":"MIT","kind":"human_freehand","licenseUrl":"https://github.com/cogtoolslab/visual_abstractions_benchmarking_public2023/blob/main/LICENSE","datasetLicenseEvidence":"https://papers.neurips.cc/paper_files/paper/2023/file/d43621ff2dfe39d298dcd4a41937c912-Supplemental-Datasets_and_Benchmarks.pdf","licenseText":"MIT License\n\n\nCopyright (c) 2019 Cognitive Tools Lab\n\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n","sourceFile":"https://seva-benchmark.s3.us-west-2.amazonaws.com/neurips_DB_human_clean.csv","method":"Retained original human cursor traces, globally deduplicated; excluded generated drawings. No reference photographs or personal fields are redistributed.","sourceSha256":"2107e92598a3d2947441cbca1a3530cca1a8c43c5b545370d1d342988426f293","sourceBytes":1244217844},"tuberlin":{"name":"TU Berlin Sketch Dataset","creator":"Mathias Eitz, James Hays, Marc Alexa","url":"https://faculty.cc.gatech.edu/~hays/papers/Classify_sketch.pdf","licenseEvidence":"https://web.archive.org/web/20250615112619/http://cybertron.cg.tu-berlin.de/eitz/projects/classifysketch/","kind":"human_freehand","license":"CC BY 4.0","licenseUrl":"https://creativecommons.org/licenses/by/4.0/"}},"previousRelease":{"url":"/data/wedraw-v3/manifest.json","sha256":"b5798b1de592e99b2254cae5e3419cda13fbf9778d5a2229e308e909a1630ea2"},"architecture":"Grayscale 24×24 → 3 learned spatial convolutions with folded normalization (12,24,32 channels); average pool → 288 spatial features concatenated with 306 normalized stroke/direction features; 192-D ReLU hidden layer → L2-normalized 96-D sketch metric. Learnable class proxies only in training, 0.09 additive cosine margin, inverse temperature 17. Runtime compares the entire shared board to two train-only componentwise median templates. One projection per board; no team metadata, external inference or D1 reads for model weights. Output is a calibrated game score, not a probability.","training":{"seed":731045,"epochs":16,"fitOriginals":187140,"perEpochMethod":"Uniform cap at 145 examples per class plus sampling without replacement of synthetic 5+5 training boards; all class medians use all original fit records; draws with fewer examples train on the full fit population.","syntheticTrainingBoards":26432,"referenceCounting":"Synthetic boards, fresh originals and alternate renders are excluded from the 229745 original references.","calibrationOriginals":21305,"heldoutOriginals":21300,"freshQuickDrawOriginals":19285},"evaluation":{"binary":"A score of 50 or greater passes a class-specific cutoff derived from the single and mixed calibration sets; negative sets include other source-matched words. Matched 24-way benchmark uses identical 12-support medians without cutoffs, and reports cosine ranks.","gameBoards":"413-way test includes 4956 equal five-plus-five boards, and two additional counterfactual layouts per pair (8:2 and 2:8 source lines). Three layout strata vary source overlap. Same data-generation protocol for all encoders. No human judgments of actual played rounds are implied.","splits":"All four groups have disjoint original drawing IDs. 345 QuickDraw classes preserve first 500, next 50 and subsequent 50 valid IDs as published; the 19285 additional test examples lie beyond the first 600 recognized examples per class. SEVA group-level photo splits, TU Berlin and CommonSketch source-group splits retained. Some simple or nearly identical shapes can occur under different IDs: 2 train/test geometry hashes matched, 2 train/calibration and 3 train/fresh. No shape hashes shared among calibration/test/fresh. No claim of author-level independence.","numericalParity":"Independent game-runtime replay of all 40585 single sketches and 9096 balanced boards exactly matched all target-hit numerators and MRR to 1e-13. One off-target threshold decision differed on the all-term mixed subset and one on the fresh single-figure subset between PyTorch/NumPy and the Worker JavaScript float accumulator: 127191 vs 127192 of 2036916 and 434523 vs 434524 of 7945420. Catalog evaluation counts use the frozen Python release evaluator.","limits":"A per-word threshold trades missed targets against cross-word false positives. Heavily overlaid objects remain difficult. DINOv2-small slightly leads this model in some 24-way retrieval metrics; no global optimality claim.","largestReferenceCollections":"This version retains 413 calibrated words. It is not a 1,000-word model."},"research":[{"title":"Sketch-a-Net (Yu et al.)","url":"https://arxiv.org/abs/1501.07873","use":"Sketch-specific spatial architecture; our model is trained from original game references rather than redistributing their checkpoint."},{"title":"No Fuss Distance Metric Learning Using Proxies (Movshovitz-Attias et al., ICCV 2017)","url":"https://openaccess.thecvf.com/content_iccv_2017/html/Movshovitz-Attias_No_Fuss_Distance_ICCV_2017_paper.html","use":"Learned class proxies supervise a compact similarity embedding."},{"title":"CosFace (Wang et al., CVPR 2018)","url":"https://openaccess.thecvf.com/content_cvpr_2018/html/Wang_CosFace_Large_Margin_CVPR_2018_paper.html","use":"Adapted normalized additive cosine-margin objective to single- and dual-label sketches. The original paper studies faces, not game drawings."},{"title":"Compositional Sketch Search (Black et al., ICIP 2021)","url":"https://research.adobe.com/publication/compositional-sketch-search/","use":"Justifies testing distinct spatial placements; our scoring receives no object masks."},{"title":"Sketch Less for More (Bhunia et al., CVPR 2020)","url":"https://openaccess.thecvf.com/content_CVPR_2020/html/Bhunia_Sketch_Less_for_More_On-the-Fly_Fine-Grained_Sketch-Based_Image_Retrieval_CVPR_2020_paper.html","use":"Partial stroke retrieval motivates the low-stroke game benchmark; our task is sketch-to-sketch rather than sketch-to-photo."},{"title":"CLIP for All Things (Sain et al., CVPR 2023)","url":"https://openaccess.thecvf.com/content/CVPR2023/html/Sain_CLIP_for_All_Things_Zero-Shot_Sketch-Based_Image_Retrieval_Fine-Grained_or_CVPR_2023_paper.html","use":"Documented domain of sketch-photo matching and need for adaptation; not benchmarked here. No CLIP model weights are served."},{"title":"DINOv2 (Oquab et al.)","url":"https://arxiv.org/abs/2304.07193","license":"Apache-2.0 original weights (Meta); only measured separately, no pretrained checkpoint redistributed.","use":"Pinned frozen DINOv2-small 384-D features measured on equal support and test records; see matched-study.json."}]}