Research map / Video / Video segmentation and tracking
DeltaFlow: An Efficient Multi-frame Scene Flow Estimation Method
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeltaFlow, a lightweight multi-frame scene flow method that applies a temporal delta (subtraction, decay weighting, summation) on sparse voxel features so input size stays constant regardless of frame count, plus Category-Balanced and Instance Consistency losses.
Problem
Multi-frame scene flow estimation on LiDAR point clouds becomes computationally expensive as the number of frames grows, since existing methods concatenate or stack per-frame features, increasing input size and memory.
Approach
Replace feature concatenation or 4D stacking with a sparse temporal delta feature (decayed sum of differences between the current voxel features and each past frame's), keeping the backbone input size fixed regardless of frame count, and add a category-balanced loss and an instance consistency loss to supervision.
Key results
- On the Argoverse 2 test leaderboard, DeltaFlow (5 frames) reaches mean dynamic bucket-normalized EPE 0.113 vs 0.130 for EulerFlow and 0.145 for Flow4D (5 frames), about 22% lower than Flow4D, while running about 2x faster.
- On nuScenes validation, DeltaFlow reduces mean three-way EPE to 2.33 cm vs 3.82 cm for Flow4D (about 39% reduction).
- In cross-domain tests (Argoverse 2 to Waymo), DeltaFlow reaches 3.12 cm mean three-way EPE, better than Flow4D at 3.33 cm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DeltaFlow with 5 frames instead of Flow4D with 5 frames | Scene flow estimation on Argoverse 2 test · Argoverse 2 test | Dynamic Bucket-Normalized EPE (Mean): 0.145 -> 0.113 | Helped |
| use DeltaFlow with 5 frames instead of Flow4D with 5 frames | Scene flow estimation on nuScenes · nuScenes validation | Three-way EPE (Mean, cm): 3.82 -> 2.33 | Helped |
| use DeltaFlow instead of Flow4D | Cross-domain scene flow estimation · Argoverse 2 to Waymo | Three-way EPE (Mean, cm): 3.33 -> 3.12 | Helped |
| use DeltaFlow instead of Flow4D | Cross-domain scene flow estimation · Waymo to Argoverse 2 | Three-way EPE (Mean, cm): 4.01 -> 3.24 | Helped |
| apply temporal decay with lambda below 1 instead of no decay (ablation) | DeltaFlow · Argoverse 2 validation, 5 frames | Dynamic Bucket-Normalized EPE (Mean): 0.2024 -> 0.1901 | Helped |
| apply temporal decay with lambda below 1 instead of no decay (ablation) | DeltaFlow · Argoverse 2 validation, 10 frames | Dynamic Bucket-Normalized EPE (Mean): 0.1967 -> 0.1901 | Helped |
| add Category-Balanced Loss to motion-based loss (ablation) | DeltaFlow · Argoverse 2 validation, 5 frames | Dynamic Bucket-Normalized EPE (Mean): 0.2094 -> 0.1962 | Helped |
| add Category-Balanced Loss to motion-based loss (ablation) | DeltaFlow · Argoverse 2 validation, 5 frames | Pedestrian Dynamic Bucket-Normalized EPE: 0.2398 -> 0.2148 | Helped |
| add Category-Balanced Loss to motion-based loss (ablation) | DeltaFlow · Argoverse 2 validation, 5 frames | CAR Dynamic Bucket-Normalized EPE: 0.1504 -> 0.1511 | Hurt |
| add Instance Consistency Loss to motion-based loss (ablation) | DeltaFlow · Argoverse 2 validation, 5 frames | Dynamic Bucket-Normalized EPE (Mean): 0.2094 -> 0.1971 | Helped |
| add Instance Consistency Loss to motion-based loss (ablation) | DeltaFlow · Argoverse 2 validation, 5 frames | Three-way EPE (Mean, cm): 3.32 -> 3.38 | Hurt |
| remove motion-based loss and use Category-Balanced Loss plus Instance Consistency Loss (ablation) | DeltaFlow · Argoverse 2 validation, 5 frames | Three-way EPE (Mean, cm): 3.26 -> 3.93 | Hurt |
Surprising findings
- Using only the category-balanced and instance consistency losses (without the motion-based loss) gives the best dynamic bucket-normalized error (0.188) but sharply increases background static three-way EPE (0.70 to 2.69 cm).
- Performance of both DeltaFlow and Flow4D does not keep improving from 10 to 15 input frames, suggesting older frames may add noise.
Limitations
- Relies on ground-truth annotations for supervised training; no self-supervised variant is provided.
- Gains plateau or decline at 15 frames, indicating long-range temporal context is not fully exploited.
- Loss weights (category weights and speed weights) were set experimentally, without a principled selection.
Experimental setup
Datasets: Argoverse 2 Waymo nuScenes
Benchmarks: Argoverse 2 scene flow leaderboard Waymo nuScenes
Metrics: Dynamic Bucket-Normalized EPE Three-way EPE runtime
Baselines: SeFlow ICP Flow ZeroFlow FastNSF NSFP Floxels EulerFlow FastFlow3D TrackFlow DeFlow SSF Flow4D
Builds on: DeFlow Flow4D MinkowskiNet PointPillars FastFlow3D
Closest papers by meaning
- MESSENGER: Memory-Enhanced Sequential Scene Flow Estimation via Autoregressive Next-Frame Forecasting (NeurIPS 2026 · similarity 0.89)
- ZeroFlow: Scalable Scene Flow via Distillation (ICLR 2024 · similarity 0.88)
- Semantic Flow: Learning Semantic Fields of Dynamic Scenes from Monocular Videos (ICLR 2024 · similarity 0.88)
- Zero-Shot Monocular Scene Flow Estimation in the Wild (CVPR 2025 · similarity 0.87)
- DifFlow3D: Toward Robust Uncertainty-Aware Scene Flow Estimation with Iterative Diffusion-Based Refinement (CVPR 2024 · similarity 0.87)
- Flow-NeRF: Joint Learning of Geometry, Poses, and Dense Flow within Unified Neural Representations (CVPR 2025 · similarity 0.87)
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.