Research map / Video / Video segmentation and tracking
Midway Network: Learning Representations for Recognition and Motion from Latent Dynamics
Christopher Hoang, Mengye Ren
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 latent dynamics modeling to base model (ablation) | Midway · BDD100K linear | mIoU: 28.3 -> 30.4 | Improved |
| add latent dynamics modeling to base model (ablation) | Midway · MPI-Sintel | EPE: 6.2 -> 4.4 | Improved |
| add motion latent refinement (on backward + multi-level) (ablation) | Midway · MPI-Sintel | EPE: 5.2 -> 3.9 | Improved |
| add gating units to forward model (ablation) | Midway · BDD100K linear | mIoU: 31.1 -> 31.5 | Improved |
| add gating units to forward model (ablation) | Midway · MPI-Sintel | EPE: 3.9 -> 4.1 | Mixed |
| add backward network to latent dynamics (ablation) | Midway · BDD100K linear | mIoU: 30.4 -> 30.0 | No clear effect |
| remove backward network from full model (ablation) | Midway · MPI-Sintel | EPE: 4.1 -> 3.7 | Mixed |
| remove multi-level from full model (ablation) | Midway · MPI-Sintel | EPE: 4.1 -> 5.2 | Hurt |
| initialize dynamics models from pretraining instead of random (ablation) | Midway · FlyingThings EPE (c) | EPE: 16.6 -> 7.3 | Improved |
| longer pretraining 100 -> 300 epochs on WT-Venice | Midway · ADE20K | mIoU: 33.4 -> 36.9 | Improved |
| scale ViT-S to ViT-B | Midway · BDD100K linear | mIoU: 39.7 -> 48.2 | Improved |
| use Midway vs CroCo v2 (ViT-B) on optical flow | CroCo v2 · FlyingThings EPE (c) | EPE: 6.1 -> 7.0 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- PooDLe — PooDLe🐩: Pooled and dense self-supervised learning from naturalistic videos (ICLR 2025) · better on 1 of 2 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
- Visual Interaction Networks: Learning a Physics Simulator from Video (NeurIPS 2017 · similarity 0.88)
- OBJECT DYNAMICS DISTILLATION FOR SCENE DECOMPOSITION AND REPRESENTATION (ICLR 2022 · similarity 0.88)
- Learning Long-term Visual Dynamics with Region Proposal Interaction Networks (ICLR 2021 · similarity 0.88)
- DynamoNet: Dynamic Action and Motion Network (ICCV 2019 · similarity 0.87)
- Learning Motion in Feature Space: Locally-Consistent Deformable Convolution Networks for Fine-Grained Action Detection (ICCV 2019 · similarity 0.87)
- PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning (ICML 2025 · similarity 0.86)
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.
