Happy to find that I've been selected as an Outstanding Reviewer for CVPR 2025!

📢 New paper CVPR 25!
Can meshes capture fuzzy geometry? Volumetric Surfaces uses adaptive textured shells to model hair, fur without the splatting / volume overhead. It’s fast, looks great, and runs in real time even on budget phones.
🔗 autonomousvision.github.io/volsurfs/
📄 arxiv.org/pdf/2409.02482

⏰ Heads up! The deadline for two #CVPR2025 Autonomous Grand Challenge tracks is May 10th, 2025:

1️⃣ NAVSIM v2 Challenge: huggingface.co/spaces/AGC20...

2️⃣ World Model Challenge by 1X: huggingface.co/spaces/1x-te...

Introducing CaRL: Learning Scalable Planning Policies with Simple Rewards
We show how simple rewards enable scaling up PPO for planning.
CaRL outperforms all prior learning-based approaches on nuPlan Val14 and CARLA longest6 v2, using less inference compute.
arxiv.org/abs/2504.17838

Loft🆙 Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models. We achieve SotA upsampling results for DINOv2. Paper and code:
andrehuang.github.io/loftup-site/

Sharing another video showing how LoftUp significantly improves DINOv2 features! Works like a charm!

Try it out:
Code: github.com/andrehuang/l...
Paper: arxiv.org/abs/2504.14032

Excited to introduce LoftUp!

A strong (than ever) and lightweight feature upsampler for vision encoders that can boost performance on dense prediction tasks by 20%–100%!

Easy to plug into models like DINOv2, CLIP, SigLIP — simple design, big gains. Try it out!

github.com/andrehuang/l...

How much 3D do visual foundation models (VFMs) know?

Previous work requires 3D data for probing → expensive to collect!

#Feat2GS @cvprconference.bsky.social 2025 - our idea is to read out 3D Gaussains from VFMs features, thus probe 3D with novel view synthesis.

🔗Page: fanegg.github.io/Feat2GS

🦣Easi3R: 4D Reconstruction Without Training!

Limited 4D datasets? Take it easy.

#Easi3R adapts #DUSt3R for 4D reconstruction by disentangling and repurposing its attention maps → make 4D reconstruction easier than ever!

🔗Page: easi3r.github.io

🐎 Centaur, our first foray into test-time training for end-to-end driving. No retraining needed, just plug-and-play at deployment given a trained model. Also, theoretically nearly no overhead in latency with some clever use of buffers. Surprising how effective this is! arxiv.org/abs/2503.11650

🚀 Names matter! We show that better class names in open-vocabulary segmentation benchmarks greatly improve dataset quality and boost model performance. RENOVATE your dataset labels with our automatic framework! #AI #ComputerVision #NeurIPS24
andrehuang.github.io/renovate/

Synchronization is ubiquitous in nature and a key mechanism for information processing in the brain. We introduce AKOrN as a dynamical alternative to threshold units, which can be combined with MLPs, CNNs or Transformers. ICLR'25 Oral. Project page: takerum.github.io/akorn_projec...

This week we had our winter retreat jointly with Daniel Cremer's group in Montafon, Austria. 46 talks, 100 Km of slopes and night sledding with some occasionally lost and found. It has been fun!