[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-detail-137722-en":59,"doc-seo-137722-105":81},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":5,"data":60},{"doc_id":61,"user_id":62,"nickname":63,"user_avatar":64,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":66,"doc_content":67,"file_id":68,"file_url":69,"file_type":70,"file_size":71,"view_count":19,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":72,"language":73,"language_code":74,"site_id":75,"html_lang":74,"table_of_contents":76,"faqs":77,"seo_title":78,"seo_description":66,"update_tm":79,"read_time":80},137722,1374402739827,"Nguyễn Văn Học","https://ap-avatar.wpscdn.com/avatar/14000c97e7351f1a627?x-image-process=image/resize,m_fixed,w_180,h_180&k=1787885694763230660","BigDatasetGAN - Synthesizing ImageNet with Pixel-wise Annotations","Pixel-wise image annotation is expensive and time-consuming, motivating generative alternatives such as DatasetGAN. BigDatasetGAN scales this idea to ImageNet by using BigGAN class-conditional generation and manually annotating only 5 images per class across 1,000 classes. A feature segmentation architecture turns these GANs into labeled dataset generators, and VQGAN can provide similar functionality using already annotated data. The work builds a new ImageNet benchmark and shows large gains for dense prediction pre-training and downstream segmentation, detection, and chest X-ray tasks.","BigDatasetGAN: Synthesizing ImageNet with Pixel-wise Annotations  \nDaiqing Li 1 Huan Ling 1 ;2 ;3 Seung Wook Kim 1 ;2 ;3  \nKarsten Kreis 1 Adela Barriuso Sanja Fidler1 ;2 ;3 Antonio Torralba4  \n1NVIDIA 2University of Toronto 3Vector Institute 4MIT fdaiqingl,huling,seungwookk,kkreis,[sfidler](sfidlerg@nvidia.com)[g](sfidlerg@nvidia.com)[@nvidia.com](sfidlerg@nvidia.com) , [torralba@mit.edu](torralba@mit.edu)  \nFigure 1 . Our synthesized pixel-wise labeled ImageNet dataset. We sample both images and masks for each of the 1k ImageNet classes.  \nAbstract  \nAnnotating images with pixel-wise labels is a timeconsuming and costly process. Recently, DatasetGAN [82] showcased a promising alternative – to synthesize a large labeled dataset via a generative adversarial network (GAN) by exploiting a small set of manually labeled, GANgenerated images. Here, we scale DatasetGAN to ImageNet scale of class diversity. We take image samples from the class-conditional generative model BigGAN [6] trained on ImageNet, and manually annotate only 5 images per class, for all 1k classes. By training an effective feature segmentation architecture on top of BigGAN, we turn BigGAN into a labeled dataset generator. We further show that VQGAN [19] can similarly serve as a dataset generator, leveraging the already annotated data. We create a new ImageNet benchmark by labeling an additional set of real images and evaluate segmentation performance in a variety of settings. Through an extensive ablation study, we show big gains in leveraging a large generated dataset to train different supervised and self-supervised backbone modelson pixel-wise tasks. Furthermore, we demonstrate that using our synthesized datasets for pre-training leads to improvements over standard ImageNet pre-training on several downstream datasets, such as PASCAL-VOC, MS-COCO, Cityscapes and chest X-ray, as well as tasks (detection, segmentation). Our benchmark will be made public and maintain a leaderboard for this challenging task. Project Page:  \n[https://nv-tlabs.github.io/big-datasetgan/](https://nv-tlabs.github.io/big-datasetgan/)  \n1. Introduction  \nThe ImageNet dataset [68] has served as a cornerstone of modern computer vision and deep learning. It is used asa testbed for innovating in the domain of large-scale classification, and has enabled incredible advancements over the years [16, 30, 44, 70] . Importantly, it has also been com-  \nmonly used for pre-training backbone models, with either supervised or recently self-supervised pre-training, leading to almost guaranteed performance gains on a plethora of downstream datasets and tasks [25, 81] . ImageNet contains a million images with 1000-way classification labels. This huge class diversity is what makes pre-trained networks generalize well to a variety of downstream applications.  \nIn this paper, we aim to enhance ImageNet with pixelwise labels, to enable large-scale multi-class segmentation challenges and offer opportunities for new pre-training strategies for dense downstream prediction tasks. However, instead of manually labeling masks for 1M images, which is time-consuming and costly, we instead synthesize high quality labeled data at a fraction of the cost.  \nWe build on top of DatasetGAN [82], which introduced a simple idea: to manually annotate a very small set of GANgenerated images with pixel-wise labels, and add a shallow segmentation branch on top of the GAN's feature maps which is trained on this small dataset. It was shown that the generator's feature maps are incredibly powerful and semantically meaningful, and allow the segmentation branch to produce very accurate labels for new random samples from the GAN. This means that the GAN is successfully re-purposed into a dataset generator, producing samples in the form of images and their pixel-wise labels. The authors showed that synthesizing a large dataset and using it to train downstream segmentation networks leads to extremely high performance at only a fraction of th","cbCaie4U0swB4WnY","https://ap.wps.com/l/cbCaie4U0swB4WnY","pdf",59391506,21,"English","en",105,"# Introduction\n## Motivation: pixel-wise labeling cost\n## DatasetGAN scaling to ImageNet\n## BigDatasetGAN pipeline overview\n# Related work","[{\"question\":\"What problem does BigDatasetGAN address?\",\"answer\":\"It addresses the high cost and time required to create pixel-wise labels for large image datasets like ImageNet.\"},{\"question\":\"How does BigDatasetGAN synthesize a labeled ImageNet-scale dataset?\",\"answer\":\"It samples images per ImageNet class from BigGAN, manually annotates only a small number per class, and trains a segmentation architecture on GAN features so the GANs generate labeled data.\"},{\"question\":\"What improvements does the benchmark and pre-training show?\",\"answer\":\"Using synthesized datasets improves segmentation, detection, and related downstream performance compared with standard ImageNet pre-training, and the benchmark is hosted with a leaderboard.\"}]","BigDatasetGAN - Synthesizing ImageNet with Pixel-wise Annotations | PDF",1787438958,53,{"code":4,"msg":82,"data":83},"ok",{"site_id":75,"language":74,"slug":84,"title":65,"keywords":85,"description":66,"schema_data":86,"social_meta":135,"head_meta":137,"extra_data":139,"updated_unix":79},"bigdatasetgan-synthesizing-imagenet-with-pixel-wise-annotations","",{"@graph":87,"@context":134},[88,102,117],{"@type":89,"itemListElement":90},"BreadcrumbList",[91,95,97,100],{"item":92,"name":93,"@type":94,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":96,"name":9,"@type":94,"position":14},"https://docshare.wps.com/document/",{"item":98,"name":40,"@type":94,"position":99},"https://docshare.wps.com/document/research-report/",3,{"item":101,"name":65,"@type":94,"position":19},"https://docshare.wps.com/document/bigdatasetgan-synthesizing-imagenet-with-pixel-wise-annotations/137722/",{"url":101,"name":65,"@type":103,"author":104,"headline":65,"publisher":106,"fileFormat":109,"inLanguage":74,"description":66,"dateModified":110,"datePublished":111,"encodingFormat":109,"isAccessibleForFree":112,"interactionStatistic":113},"DigitalDocument",{"name":63,"@type":105},"Person",{"url":92,"name":107,"@type":108},"DocShare","Organization","application/pdf","2026-09-03","2026-08-22",true,{"@type":114,"interactionType":115,"userInteractionCount":19},"InteractionCounter",{"@type":116},"ViewAction",{"@type":118,"mainEntity":119},"FAQPage",[120,126,130],{"name":121,"@type":122,"acceptedAnswer":123},"What problem does BigDatasetGAN address?","Question",{"text":124,"@type":125},"It addresses the high cost and time required to create pixel-wise labels for large image datasets like ImageNet.","Answer",{"name":127,"@type":122,"acceptedAnswer":128},"How does BigDatasetGAN synthesize a labeled ImageNet-scale dataset?",{"text":129,"@type":125},"It samples images per ImageNet class from BigGAN, manually annotates only a small number per class, and trains a segmentation architecture on GAN features so the GANs generate labeled data.",{"name":131,"@type":122,"acceptedAnswer":132},"What improvements does the benchmark and pre-training show?",{"text":133,"@type":125},"Using synthesized datasets improves segmentation, detection, and related downstream performance compared with standard ImageNet pre-training, and the benchmark is hosted with a leaderboard.","https://schema.org",{"og:url":101,"og:type":136,"og:title":65,"og:site_name":107,"og:description":66},"article",{"robots":138,"canonical":101},"index,follow",{"doc_id":61,"site_id":75}]