Research map / Point clouds / Point cloud classification
Generative Point Tracking and Forecasting
Xuanchen Lu, Ang Cao, Chao Feng, Andrew Owens
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 |
|---|---|---|---|
| use diffusion (flow matching) loss instead of L1 regression (ablation) | point tracking model · Tracking ablation | AJ: 43.3 -> 53.7 | Improved |
| use diffusion loss instead of L2 regression (ablation) | point tracking model · Tracking ablation | AJ: 30.2 -> 53.7 | Improved |
| add factorization of query points during training (ablation) | point tracking model · Tracking ablation | AJ: 36.6 -> 53.7 | Improved |
| add tuned logit-normal noise schedule (ablation) | point tracking model · Tracking ablation | AJ: 51.6 -> 53.7 | Improved |
| use stride 1 instead of larger sliding window stride (ablation) | forecasting model · Kubric forecasting | FVMD (Scene): 23361 -> 17786 (stride 8 vs 1) | Improved |
| lower context noise level (0.15 to 0.02) (ablation) | forecasting model · Kubric forecasting | FVMD (Scene): 21552 -> 17786 | Improved |
| lower context noise 0.02 to 0.05 (ablation) | forecasting model · Kubric forecasting | Best of K: 95.6 -> 92.7 | Mixed |
| increase tracking context frames from 1 to 8 before forecasting (ablation) | unified model · Kubric and Physics101 forecasting | forecast error: Physics101 18.99 -> 11.42 | Improved |
| train jointly on tracking and forecasting (unified) instead of tracking only (ablation) | unified model · TAP-Vid DAVIS | AJ: 61.7 -> 59.9 | Hurt |
| unified training with in-domain motion prior under synthetic occlusion (ablation) | unified model · Physics101 tracking with occlusion | delta_occ_avg: 68.8 -> 82.2 | Improved |
| unified training with out-of-domain motion prior (Kubric-trained, DAVIS eval) (ablation) | unified model · DAVIS tracking with occlusion | delta_vis_avg: 75.6 -> 73.8 | Hurt |
| point-space flow matching model vs latent diffusion forecaster | Ours (Forecasting Only) · Kubric MOVi-A forecasting | Best of K: 127.0 -> 95.6 | Improved |
Closest papers by meaning
- Track-On: Transformer-based Online Point Tracking with Memory (ICLR 2025 · similarity 0.85)
- Track4Gen: Teaching Video Diffusion Models to Track Points Improves Video Generation (CVPR 2025 · similarity 0.85)
- MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.84)
- E-Motion: Future Motion Simulation via Event Sequence Diffusion (NeurIPS 2024 · similarity 0.84)
- Tracktention: Leveraging Point Tracking to Attend Videos Faster and Better (CVPR 2025 · similarity 0.84)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.83)
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.
