Research map / Video / Video segmentation and tracking
Unbiased Gradient Estimation for Event Binning via Functional Backpropagation
Jinze Chen, Wei Zhai, Han Han, Tiankai Ma, Yang Cao, Bin Li, Zheng-Jun Zha
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 |
|---|---|---|---|
| synthesize weak-derivative gradients via functional backprop | contrast-maximization optimization · ECD angular velocity estimation | RMS error: -10.3% RMS, 1.66x faster | Improved |
| synthesize weak-derivative gradients via functional backprop | contrast-maximization optimization · ECD linear velocity estimation | RMS error: +3.9% RMS error, 1.48x faster | Mixed |
| synthesize weak-derivative gradients via functional backprop | contrast-maximization optimization · ECD motion estimation overall | RMS error: -3.2% RMS, 1.57x speedup | Improved |
| use FBP instead of STE surrogate (ablation) | FBP · shapes_rotation Var | RMS error (deg/s): 96.89 -> 15.89 | Improved |
| use FBP instead of sigmoid surrogate (ablation) | FBP · boxes_rotation Var | RMS error (deg/s): 13.94 -> 12.44 | Improved |
| use linear instead of bicubic/Lanczos reconstruction kernel (ablation) | FBP · ECD rotation, Var and LL | RMS error and time: accuracy within ~0.5 deg/s, faster time e.g. 7.44 vs 8.56 ms | Mixed |
| replace linear binning with rect kernel using FBP gradients | MotionPriorCMax · DSEC optical flow | EPE: -9.4% EPE | Improved |
| replace linear binning with rect kernel using FBP gradients | CMax-SLAM · ECD first 30s | Abs RMS error: -5.1% Abs RMS | Improved |
| apply FBP to log-likelihood score with linear binning kernel | bias analysis · dynamic_rotation LL score | gradient bias | Hurt |
| apply FBP gradient computation | JVP of binning · computation time, GPU Ne=20000 | time (us): Linear JVP 28.1 -> FuncLinear JVP 66.8 | Hurt |
Closest papers by meaning
- Secrets of Event-Based Optical Flow, Depth and Ego-Motion Estimation by Contrast Maximization (TPAMI 2024 · similarity 0.87)
- E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization (NeurIPS 2025 · similarity 0.86)
- Unsupervised Event-Based Learning of Optical Flow, Depth, and Egomotion (CVPR 2019 · similarity 0.86)
- Discrete Time Convolution for Fast Event-Based Stereo (CVPR 2022 · similarity 0.86)
- EventZoom: Learning To Denoise and Super Resolve Neuromorphic Events (CVPR 2021 · similarity 0.85)
- Spatio-Temporal Recurrent Networks for Event-Based Optical Flow Estimation (AAAI 2022 · similarity 0.85)
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.
