Research map / Point clouds / Point cloud classification
A Simple and Effective Point-based Network for Event Camera 6-DOFs Pose Relocalization
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.
Introduces PEPNet, a lightweight point-cloud-style network that regresses 6-DOF event camera poses directly from raw event clouds using hierarchical point feature extraction, temporal attention aggregation, and an attentive Bi-LSTM.
Problem
Camera pose relocalization (CPR) from event cameras: existing event-based CPR networks convert events into frame-like representations, losing fine-grained temporal information, and are computationally heavy for edge devices.
Approach
Treat events as a pseudo-3D point cloud (x, y, normalized t) processed without permutation invariance in a hierarchical PointNet++-style network, replacing max pooling with temporal attention aggregation and adding an attentive Bi-LSTM to capture explicit temporal dependencies.
Key results
- PEPNet reports average random-split IJRR error of 0.013m and 0.904° versus 0.019m and 1.591° for CNN-LSTM, described as a 38% improvement.
- PEPNet-tiny uses 0.064M parameters (about 0.5% of CNN-LSTM's 12.63M) and reaches 0.019m and 1.306° average error on IJRR random split.
- On the IJRR novel split, PEPNet achieves average 0.029m and 2.13° versus 0.035m and 2.23° for AECRN, with 6.7ms inference time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| process raw events as a point cloud with PEPNet | Event camera pose relocalization · IJRR random split, Average | translation error: 0.019m -> 0.013m | Helped |
| process raw events as a point cloud with PEPNet | Event camera pose relocalization · IJRR novel split, Average | rotation error: 2.23◦ -> 2.13◦ | Helped |
| process raw events as a point cloud with PEPNet | Event camera pose relocalization · M3ED random split, Average | rotation error: 2.197◦ -> 1.04◦ | Helped |
| add LSTM to the hierarchy-only model (ablation) | PEPNet · IJRR random split, shape translation | T+R: 3.04 -> 2.77 | Helped |
| add Bi-LSTM rather than LSTM (ablation) | PEPNet · IJRR random split, shape translation | T+R: 2.85 -> 2.12 | Helped |
| add temporal attention aggregation instead of Max aggregation (ablation) | PEPNet · IJRR random split, shape translation | T+R: 2.82 -> 2.12 | Helped |
| change loss coefficients from α=0.5, β=0.5 to α=0.25, β=0.75 (ablation) | PEPNettiny · IJRR random split, shape translation | translation error: 0.0302m -> 0.0359m | Hurt |
| change loss coefficients from α=0.5, β=0.5 to α=0.75, β=0.25 (ablation) | PEPNettiny · IJRR random split, shape rotation | rotation error: 2.888◦ -> 3.36◦ | Hurt |
Surprising findings
- PEPNet does not require pre-trained weights, unlike SP-LSTM and AECRN, but shows more overfitting on the novel split.
Limitations
- Overfitting is more pronounced on the novel split due to train/test distribution differences and limited data.
- Inference time is dominated by grouping and sampling, and real-time edge deployment is not demonstrated beyond server timing.
- Evaluated only on IJRR and M3ED event CPR datasets with scene-specific training.
Experimental setup
Datasets: IJRR event-based CPR dataset M3ED
Benchmarks: IJRR M3ED
Metrics: median translation error (m) median rotation error (degrees) MSE parameters FLOPs inference time
Baselines: PoseNet Bayesian PoseNet Pairwise-CNN LSTM-Pose SP-LSTM CNN-LSTM AECRN DSAC*
Builds on: PointNet++ LSTM
Closest papers by meaning
- SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition (ICLR 2024 · similarity 0.86)
- State Space Models for Event Cameras (CVPR 2024 · similarity 0.85)
- Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization (CVPR 2025 · similarity 0.84)
- Continuous Pose for Monocular Cameras in Neural Implicit Representation (CVPR 2024 · similarity 0.84)
- PASS: Path-selective State Space Model for Event-based Recognition (NeurIPS 2025 · similarity 0.84)
- GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting (ICLR 2025 · 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.