Nikita Araslanov

📢Applications are now open for PhD visiting students @bristolcs @BristolUni in 2026 - DL 29 Jan. Would you like to work with any of the Faculty working in Machine Learning and Computer Vision #mavi as part of our summer of research @ Bristol program? @MaVi.html" target="_blank" rel="nofollow noopener">uob-mavi.github.io/Summer@MaVi.ht…

@CSProfKGD @Michael_J_Black Thank you @CSProfKGD ☺️ That’s the spiral of research, revisiting some old ideas in the new context…

Cool idea @neekans!❤️ Reminded me of some of @Michael_J_Black’s work on steerable flow fields, eg MJ Black, Y Yacoob, AD Jepson, DJ Fleet, Learning parameterized models of image motion, CVPR, 1997 💪

Afternoon e-reading

@CSProfKGD Thanks for reading. I will improve 😉

Repeat after me: This is NOT a paper contribution, it’s an expected component!

📢 NeurIPS 2025 Spotlight 📢 Can we embed motion into image representations? Trained on videos, FlowFeat embeds optical flow into pixel-level representations (up to a linear transform), which results in sharp feature grids, especially for dynamic objects. We demonstrate benefits for ⭐️video object segmentation; ⭐️semantic segmentation; ⭐️and monocular depth. Paper: arxiv.org/abs/2511.07696 Project website: tum-vision.github.io/flowfeat Code and models: github.com/tum-vision/flo… Joint work with Anna Sonnweber and Daniel Cremers @tumcvg and @MunichCenterML. Come by our poster @NeurIPSConf on Thursday (Exhibit Hall C,D,E #4816)!

FlowFeat distills optical flow networks into pixel-level task-agnostic representations. FlowFeat provides versatile pixel-level features. Using motion-driven embedding statistics, it achieves high spatial precision and temporal consistency Paper Title: FlowFeat: Pixel-Dense Embedding of Motion Profiles Project: tum-vision.github.io/flowfeat Link: arxiv.org/abs/2511.07696 #Video #Motion #MotionGraphics #AI #AIイラスト

Paper: openreview.net/pdf?id=q9r2DzP…

#ICCV2025 Spotlight talk (SP4V Workshop) Training on videos should yield more 3D-aware models than images — but it doesn’t! Presenting The Diashow Paradox in our talk, 16:15 @ 323A. With Tien Duc Nguyen (@hoaquin10), Anna Sonnweber, Mark Weber, Daniel Cremers (@tumcvg)

@tumcvg goes #ICCV2025 in Hawaii! 🛫🌋 We are very proud of our students who will present five papers (+ 1 workshop) during the conference! In particular, check out Back-on-track, which is an award candidate. (Congrats @wrchen530!)

We are presenting 3 papers at #CVPR2025! Scene-Centric Unsupervised Panoptic Segmentation by @olvr_hhn, @ChristophR1996, @neekans, Daniel Cremers, Christian Rupprecht, and @stefanroth Sunday, 8:30 AM, ExHall D, Poster 330 Project Page: visinf.github.io/cups

The implication is quite exciting: unsupervised classifiers. We can assign semantic labels to visual data without any paired data, at least for some semantic concepts. #CVPR2025

Can we match vision and language representations without any supervision or paired data? Surprisingly, yes!  Our #CVPR2025 paper with @neekans and Daniel Cremers shows that the pairwise distances in both modalities are often enough to find correspondences. ⬇️1/4

@zhifan_zhu @ducha_aiki @felixwimbauer Yes, it works quite (and surprisingly) well. Feel free to reach out to @wrchen530 to learn more 😉

@ducha_aiki @felixwimbauer @neekans Do they really decouple the static points from dynamic cleanly? In Eq-(1) they have `m` a floating number in [0, 1], I'm curious what those m=0.5 points look like.

Back on Track: Bundle Adjustment for Dynamic Scene Reconstruction Weirong Chen, Ganlin Zhang, @felixwimbauer , Rui Wang, @neekans Andrea Vedaldi, Daniel Cremers tl;dr even for non-rigid SfM you can do BA on static parts -> improves everything. arxiv.org/abs/2504.14516

@SattlerTorsten @ducha_aiki @felixwimbauer Still more work to be done: masking out dynamic points yet works slightly better for VO. However, it’s surprising (at least to me) that a point tracker can disentangle the dynamic and static components from the observed motion (which can be really complex).

@SattlerTorsten @ducha_aiki @felixwimbauer You’re quite right about previous work! However, we don’t filter out dynamic points. Instead, we remove the dynamic motion component of those points, keeping the static one (due to camera motion) and pass all points to BA.

Back on Track: Bundle Adjustment for Dynamic Scene Reconstruction @wrchen530, @zhang_ganlin, @felixwimbauer, Rui Wang, @neekans, Andrea Vedaldi, Daniel Cremers tl;dr: learning-based 3D point tracker decouples camera and object-based motion arxiv.org/abs/2504.14516

📢 #CVPR2025: Scene-Centric Unsupervised Panoptic Segmentation 🔥 We present CUPS, the first unsupervised panoptic segmentation method trained directly on scene-centric imagery. Using self-supervised features, depth & motion, we achieve SotA results! 🌎 visinf.github.io/cups

📣 #CVPR2025 (Highlight): Scene-Centric Unsupervised Panoptic Segmentation Check out our recent CVG paper on unsupervised panoptic segmentation!🚀

We are thrilled that our group has twelve papers accepted at #CVPR2025! 🚀 Congratulations to all of our students for this great achievement! 🎉 For more details, check out: cvg.cit.tum.de

My group is looking for motivated PhD students that want to work on the future of digital humans. Within the ERC project 'LeMo: Learning Digital Humans in Motion' there are two open positions: career.tu-darmstadt.de/HPv3.Jobs/TU-D… career.tu-darmstadt.de/HPv3.Jobs/TU-D…

@andreasgeiger0 @tumcvg Hope you're having a great time!