[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83430-en":3,"doc-seo-83430-105":29,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},83430,7971461741311,"Ophelia","https://ap-avatar.wpscdn.com/avatar/74000253aff267980c6?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779345379180704826",8,"Research & Report","ZipDepth Bringing Lightweight Zero-Shot Monocular Depth Anywhere on Any Device","Monocular depth estimation improves with foundation models that support robust zero-shot generalization, but their size and compute needs make them impractical for embedded and mobile deployment. Existing lightweight methods often rely on single-domain self-supervision and can fail silently under domain shift. ZipDepth introduces a compact monocular depth network using an efficient reparameterizable encoder-decoder and large-scale knowledge distillation from a foundation model, achieving real-time inference with a strong accuracy-efficiency trade-off across multiple benchmarks.","arXiv :2607 .0877 1v 1 [ cs .CV] 9 Jul 2026  \nZipDepth: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device  \nFabio Tosi, Luca Bartolomei, Matteo Poggi, and Stefano Mattoccia  \nUniversity of Bologna, Bologna, Italy  \n{fabio.tosi5, luca.bartolomei5, m.poggi, [stefano.mattoccia}@unibo.it](stefano.mattoccia}@unibo.it)[ ](stefano.mattoccia}@unibo.it)[https://zipdepth.github.io/](https://zipdepth.github.io/)  \nFig. 1: ZipDepth generalizes zero-shot across diverse and challenging scenes. Top row: input RGB; bottom row: predicted depth at real-time rates even on a 15 W Jetson Orin NX (34 FPS with PyTorch Eager FP32, up to 77 FPS with TensorRT FP16) .  \nAbstract. Monocular depth estimation has seen remarkable progress through foundation models achieving robust zero-shot generalization, yet their computational demands place them far beyond the reach of embedded and mobile platforms. Lightweight alternatives exist, but have been developed almost exclusively within single-domain, self-supervised paradigms, failing silently under domain shift. We present ZipDepth, a compact monocular depth network that bridges this gap by combining an efficient reparameterizable encoder-decoder with large-scale knowledge distillation from a foundation model over a large multi-domain training set. Comprising just 6.1M parameters, ZipDepth runs at real-time rates from server GPUs to power-constrained devices, achieving the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models across five benchmarks, taking a significant step towards the accuracy of foundation models with 50× more parameters.  \n1 Introduction  \nMonocular depth estimation has long been regarded as one of the most ill-posed problems in computer vision: a single 2D image can correspond to infinitely many 3D scenes, leaving geometric reasoning alone insufficient to resolve the ambiguity. Yet it remains a critical capability for a wide range of real-world  \n2 F. Tosi et al.  \nMean AbsRel Error (%) → Lower is better  \n20  \n18  \n16  \n14  \n12  \n10  \n8  \n| Embedded Regime (\u003C 2:5J / frame)\u003Cbr>\u003Cbr>\u003Cbr>\u003Cbr>\u003Cbr>\u003Cbr>Better Efficiency & Accuracy |  |  |  |  Large Foundation Models\u003Cbr> Lightweight Models\u003Cbr> ZipDepth (Ours) |  |  |\n| --- | --- | --- | --- | --- | --- | --- |\n|  |  |  |  |  |  |  |\n|  |  |  |  |  |  |  |\n|  |  |  |  |  |  |  |\n|  |  |  |  |  |  |  |\n|  |  |  |  |  |  |  |\n|  |  |  |  |  |  |  |\n|  |  |  |  |  | Inference Speed\u003Cbr> 5 FPS\u003Cbr>20 FPS\u003Cbr>35 FPS\u003Cbr> |  |\n|  |  |  |  |  |  |  |\n\n    103        104        105  \nEnergy per frame (mJ) → Lower is better  \nFig. 2: Energy–Accuracy Trade-off. Zero-shot depth models on NVIDIA Jetson Orin NX (15W) . Energy per frame (FP32 inference) vs. mean AbsRel (over 5 datasets); bubble size indicates FPS. ZipDepth bridges the gap between lightweight and foundation models, tending toward their accuracy under a \u003C 400 mJ/frame budget.† denotes lightweight baselines retrained on our multi-domain training set for a fair comparison.  \nsystems, including autonomous vehicles, robotic manipulation, augmented reality, and scene reconstruction, where accurate, dense depth is required without incurring the cost or complexity of active sensing.  \nInitial learning-based approaches have transformed this problem. Supervised methods [5, 18 , 20] demonstrated that CNNs could recover plausible depth from appearance cues alone, while self-supervised formulations [22, 26 , 106] removed the dependency on ground-truth annotations by exploiting photometric consistency across views at training time. However, both paradigms were limited to specific training distributions, usually a single dataset, sensor, or scene category, while performing poorly when used outside their training domain.  \nThe recent emergence of foundation models has fundamentally changed this picture. By leveraging powerful pre-trained encoders, such as DINOv2 [48] and large Vision Transformers [16], or diffusion backbones pre-trained on billions of imag","cbCaidCSWlXLK6e6","https://ap.wps.com/l/cbCaidCSWlXLK6e6","pdf",17007502,1,23,"English","en",105,"# Introduction\n## Motivation and limitations of existing methods\n## Foundation models and deployment constraints\n## Lightweight models and domain shift problem\n## ZipDepth approach overview","[{\"question\":\"Why are foundation models for monocular depth estimation hard to deploy on mobile or embedded devices?\",\"answer\":\"They require tens of high-end GPUs to train, contain hundreds of millions of parameters, and demand very high compute per forward pass. As a result, inference speed and energy budgets on devices like Jetson Orin NX and smartphones make them impractical.\"},{\"question\":\"What issue limits many existing lightweight monocular depth models?\",\"answer\":\"They are developed mostly within a single-domain self-supervised paradigm and are trained/evaluated on datasets such as KITTI or NYUv2. This causes substantial accuracy degradation when the test domain shifts.\"},{\"question\":\"How does ZipDepth combine efficiency with zero-shot generalization?\",\"answer\":\"ZipDepth uses a compact encoder-decoder with reparameterizable efficiency and trains via large-scale knowledge distillation from a foundation model across a multi-domain dataset. This is designed to close the gap between lightweight efficiency and foundation-model breadth.\"}]",1784187623,58,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"zipdepth-bringing-lightweight-zero-shot-monocular-depth-anywhere-on-any-device","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/zipdepth-bringing-lightweight-zero-shot-monocular-depth-anywhere-on-any-device/83430/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-22","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why are foundation models for monocular depth estimation hard to deploy on mobile or embedded devices?","Question",{"text":75,"@type":76},"They require tens of high-end GPUs to train, contain hundreds of millions of parameters, and demand very high compute per forward pass. As a result, inference speed and energy budgets on devices like Jetson Orin NX and smartphones make them impractical.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What issue limits many existing lightweight monocular depth models?",{"text":80,"@type":76},"They are developed mostly within a single-domain self-supervised paradigm and are trained/evaluated on datasets such as KITTI or NYUv2. This causes substantial accuracy degradation when the test domain shifts.",{"name":82,"@type":73,"acceptedAnswer":83},"How does ZipDepth combine efficiency with zero-shot generalization?",{"text":84,"@type":76},"ZipDepth uses a compact encoder-decoder with reparameterizable efficiency and trains via large-scale knowledge distillation from a foundation model across a multi-domain dataset. This is designed to close the gap between lightweight efficiency and foundation-model breadth.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]