Research map / Video / Video segmentation and tracking
Featurising Pixels from Dynamic 3D Scenes with Linear In-Context Learners
Nikita Araslanov, Martin Sundermeyer, Hidenobu Matsuki, David Joseph Tan, Federico Tombari
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add linear in-context learning instead of ERM distillation of the same cues (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 61.1 -> 73.9 | Improved |
| add depth and flow cues on top of self-distillation (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 68.6 -> 73.9 | Improved |
| add self-distillation cue (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 72.6 -> 73.9 | Improved |
| add depth cue (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 72.6 -> 73.9 | Improved |
| add optical flow cue (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 72.5 -> 73.9 | Improved |
| add flow cue (ablation) | LILA (DINOv2-S14) · COCO-Stuff | mIoU: 59.3 -> 59.6 | No clear effect |
| add PAMR refinement of encoder features (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 71.9 -> 73.9 | Improved |
| add cropping to synthesise queries (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 72.4 -> 73.9 | Improved |
| add temporal sampling of frame pairs (ablation) | LILA (DINOv2-S14) · DAVIS-2017 | J&F (LP / k-NN): 69.3 / 72.4 -> 68.6 / 73.9 | Mixed |
| add edge (gradient matching) loss (ablation) | LILA (DINOv2-S14) · DAVIS-2017 (k-NN) | J&F: 72.9 -> 73.9 | Improved |
| train LILA on Kinetics instead of YouTube-VOS | LILA (DINOv2-L14) · DAVIS-2017 (linear probing) | J&F: 69.0 -> 69.3 | Improved |
| train LILA on Kinetics instead of YouTube-VOS | LILA (DINOv2-L14) · COCO-Stuff | mIoU: 62.8 -> 63.3 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FlowFeat — FlowFeat: Pixel-Dense Embedding of Motion Profiles (NeurIPS 2025) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning (CVPR 2021 · similarity 0.88)
- Context Encoders: Feature Learning by Inpainting (CVPR 2016 · similarity 0.87)
- In Pursuit of Pixel Supervision for Visual Pre-training (CVPR 2026 · similarity 0.87)
- UVU: Improving Multimodal Understanding via Vision-Language Unified Autoregressive Paradigm (CVPR 2026 · similarity 0.85)
- Integrating Low-Level Visual Cues for Enhanced Unsupervised Semantic Segmentation (AAAI 2025 · similarity 0.84)
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
