[1/6] Recent models like DUSt3R generalize well across viewpoints, but performance drops on aerial-ground pairs. At #CVPR2025, we propose AerialMegaDepth (, a hybrid dataset combining mesh renderings with real ground images (MegaDepth) to bridge this gap.

Will be presenting our poster shortly at 📍#59!. #CVPR2025.

Ground-aerial 3D prediction is tough! @zhiwen_fan_'s work highlights the struggle for existing models (DUSt3R/MASt3R/VGGT/etc.), indicating that much work still needs to be done in this area. At #CVPR2025, we will be presenting AerialMegaDepth (, a hybrid.

RT @anishmadan23: 🚨 The 2nd iteration of our @CVPR Foundational Few-Shot Object Detection Challenge is LIVE!. Can your model think like an….

[6/6] For more info, check out our project page, with code also released:.📄 And don’t miss the interactive web viewer for a snapshot of our data:.🌐 Work done w/ @anuragxel, @RamananDeva, Srinivasa Narasimhan, and @shubhtuls.

[5/6] Improved novel view synthesis: ZeroNVS, when fine-tuned on AerialMegaDepth, shows clear gains in both visual quality and viewpoint accuracy when synthesizing extreme aerial-to-ground viewpoint changes.

[4/6] Through a mix of real and mesh-rendered training pairs, finetuning DUSt3R/MASt3R on AerialMegaDepth boosts aerial-ground reconstruction. This enables a neat capability: merging non-overlap ground images into a shared 3D scene using aerial views as a "map"/global context.

[3/6] Using classical SfM, we co-register pseudo-synthetic aerial renderings from 3D meshes (Google Earth) with real ground images from MegaDepth. Result: AerialMegaDepth w/ 132K images in 137 scenes with poses/intrinsics/depths in a unified frame—suitable for learning 3D tasks!

[2/6] Why do models fail on aerial-ground pairs?. A key reason is the lack of co-registered training data. Datasets like MegaDepth contain mostly ground or aerial images — but not both. Finetuning DUSt3R on AerialMegaDepth boosts accuracy from <5% to ~56% (rotation error < 5°).

RT @nnsriram97: 𝐂𝐚𝐧 𝐰𝐞 𝐢𝐧𝐟𝐞𝐫 𝐚𝐧 𝐨𝐛𝐣𝐞𝐜𝐭’𝐬 𝐬𝐡𝐚𝐩𝐞 𝐛𝐲 𝐨𝐛𝐬𝐞𝐫𝐯𝐢𝐧𝐠 𝐢𝐭𝐬 𝐡𝐞𝐚𝐭 𝐟𝐥𝐨𝐰? Check out our recent ECCV paper "𝐒𝐡𝐚𝐩𝐞 𝐟𝐫𝐨𝐦 𝐇𝐞𝐚𝐭 𝐂𝐨𝐧𝐝𝐮𝐜𝐭𝐢𝐨𝐧" to….

RT @CSProfKGD: @eccvconf #ECCV2020 decisions are now up on OpenReview.

RT @jeremyphoward: "PyTorch is becoming the deep learning framework of choice at Microsoft" - Big statement from @Microsoft at @PyTorch dev….

RT @ylecun: PyTorch 1.3 is live!

RT @karpathy: To understand X I need to review my Chemistry. *2 hours pass*. To understand this part I need to review my Quantum Mechanics.….

RT @breandan: Dr. @DavidDuvenaud doing double duty deriving differential equations and demonstrating how to be a good dad. Full paper: http….

RT @chrmanning: Just a couple of years ago, I found it hard to believe that vision people were still all working with rectangular bounding….