Visual Inference Lab

Visual Inference Lab of @stefanroth at @TUDarmstadt. Research in Computer Vision and Machine Learning.

[Visual Inference Lab @visinf]

Work by: @ChristophR1996*, @olvr_hhn*, @neekans, @lealtaixe, C. Rupprecht, D. Cremers and @stefanroth

📄Paper: https://t.co/JStX3RQLEd
🌍Project Page: https://t.co/FpSVtkyiaZ
💻Code: https://t.co/D7gczjLFBH
📹Video: https://t.co/32K3HoXSMg
👁️CVPR: Friday, Poster Session 2 #333

[Visual Inference Lab @visinf]

📢 [CVPR’26] Can we learn to detect, segment, and track every object in a video without human supervision? 

Yes, we introduce VideoCUPS, the first unsupervised video panoptic segmentation (VPS) method: 1. Get pseudo-labels from monocular videos. 2. Train a VPS model on them.

[Visual Inference Lab @visinf]

When fine-tuned with just 10% of labels, VideoCUPS already matches a fully supervised model trained on all Cityscapes-VPS labels, and outperforms the DINO-initialized baseline significantly.

[Visual Inference Lab @visinf]

[6/6] Multimodal Knowledge Distillation for Egocentric Action Recognition Robust to Missing Modalities

@dustin_carrion*, M. Santos-Villafranca*, A. Perez-Yus, J. Bermudez-Cameo, J.J. Guerrero, @schaub_simone

Paper: https://t.co/nzKtI4rcMn
Project Page: https://t.co/wyhQxxq8R2

[Visual Inference Lab @visinf]

[1/6] 📢 We are in Denver at #CVPR2026 presenting 5 papers!

[Visual Inference Lab @visinf]

[5/6] MUFASA: A Multi-Layer Framework for Slot Attention

S. Bock*, L. Schüßler*, @krissingh_ , @schaub_simone , @stefanroth

Paper: https://t.co/5MWZUczdoH
Project Page: https://t.co/p5FvnQWGNo

[Visual Inference Lab @visinf]

[3/3] Project page: https://t.co/QH7qZ6Ytq6

Poster (ICRA): Thursday, 03:00 PM, P207 (Hall C - ThI2I)
Poster (CVPRW): Thursday, 10:00 AM, A2A-MML Workshop, Hall A

[Visual Inference Lab @visinf]

[1/3] Multimodal Knowledge Distillation for Egocentric Action Recognition Robust to Missing Modalities
by @dustin_carrion*, Maria Santos-Villafranca*, Alejandro Perez-Yus, Jesus Bermudez-Cameo, Jose J. Guerrero, and @schaub_simone

[Visual Inference Lab @visinf]

[2/3] KARMMA is a multimodal-to-multimodal distillation framework for egocentric action recognition that does not require modality-aligned data and supports any subset of modalities at inference. It produces a lightweight student robust to missing modalities without retraining.

[Nikita Araslanov @neekans]

In-context learning suggests that a model has learned versatile representations. What if we use in-context learning itself as a training task for visual representations?

📣 Introducing 𝗟𝗜𝗟𝗔: 𝗟𝗶𝗻𝗲𝗮𝗿 𝗜𝗻-𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 ✨ @CVPR 2026 Oral ✨

𝗟𝗜𝗟𝗔 trains on videos without manual annotation.
Key idea: An optimal linear mapping that predicts dense cues (e.g. depth, flow), estimated on one video frame, should also predict the corresponding cues of another frame from the same video.
This yields compelling results on dense vision tasks: video object segmentation, (zero-shot) semantic segmentation and surface normal estimation.

Paper, code, models and demo: https://t.co/Xn2SgskKQ8

Joint work with @ma_sundermeyer, Hidenobu Matsuki, David Joseph Tan and @fedassa (and special thanks to David and Federico for hosting my research visit at Google).

#cvpr2026 @Google @MunichCenterML @tumcvg @TU_Muenchen

[Claudia Cuttano @ClaudiaCuttano]

✨#CVPR2026 Oral ✨

A tale of a failed experiment: what if you fine-tune DINOv2 on sparse keypoints, beat every benchmark, only to discover it performs worse than the original frozen model on novel keypoints?

🚀MARCO closes this gap: a unified model for generalisable correspondences

https://t.co/vE62YiTVfd

[Gabriele Trivigno @gabTrivv]

What if a model could learn dense semantic matches from just a handful of annotated landmarks, while still generalizing to unseen keypoints and categories — and running 10× faster than diffusion-based approaches?

MARCO is selected as an Oral at #CVPR2026! A unified model for generalizable semantic correspondence, built on DINOv2⭐️

👉 Try our model: https://t.co/Zvt4QTRVJQ

[Claudia Cuttano @ClaudiaCuttano]

✨ As a first-year PhD student, I used to wonder what it must feel like to have a paper selected as an Oral at #CVPR. Today, I’m experiencing that feeling twice!

I’m beyond happy to share that both of my first-author papers have been selected as #Oral at #CVPR2026 🎉

[Gabriele Trivigno @gabTrivv]

🔥 Can in-context segmentation emerge directly from frozen DINOv3 features?

At #CVPR2026, we present INSID3: Training-Free In-Context Segmentation with DINQv3 — a collaboration between PoliTo, TU Darmstadt and TU Munich.

A training free approach that generalizes from object-level to part-level and personalized segmentation, across natural, medical, underwater, and aerial domains

Check it out: https://t.co/AaMRbfjLyn