Research map / Architectures and efficiency / Spiking networks
SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition
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 SpikePoint, an end-to-end point-cloud spiking neural network with a single-stage local/global feature extractor and an absolute-value rate-coding scheme for event-camera action recognition, trained with surrogate gradients.
Problem
Event-camera action recognition with SNNs usually converts asynchronous events into frames, losing sparsity and fine temporal detail; point-cloud networks are not directly suited to binary spike inputs, especially with negative relative coordinates.
Approach
Treat event streams as spatio-temporal point clouds processed by a single-stage SNN with identity-mapping residual blocks, and handle negative relative coordinates by taking absolute values before rate coding, with a min-coordinate branch to compensate.
Key results
- Achieves 98.74% accuracy on DVS128 Gesture with 0.58M parameters, the best among SNNs reported in the paper.
- Achieves 97.92% on Daily DVS with 0.16M parameters, and 90.6% on DVS Action with 0.16M parameters, reported as SOTA over listed methods.
- Achieves 55.6% on HMDB51-DVS (0.79M params) and 68.46% on UCF101-DVS (1.05M params), the latter best among SNNs in the table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use absolute-value processing instead of [0,1] normalization for relative coordinates (ablation) | SpikePoint grouping · Daily DVS | accuracy: 96.53% -> 97.78% | Helped |
| add minimum coordinates to the single-branch [0,1] relative-coordinate input (ablation) | SpikePoint grouping · Daily DVS | accuracy: 96.53% -> 97.36% | Helped |
| add a centroid feature branch to the absolute-relative-coordinate and minimum-coordinate input (ablation) | SpikePoint grouping · Daily DVS | accuracy: 97.63% -> 97.92% | Helped |
| use a dual-path structure instead of a single path (ablation) | SpikePoint grouping · Daily DVS | accuracy: 97.78% -> 97.92% | Helped |
| use Add instead of Concat for feature fusion (ablation) | SpikePoint grouping · Daily DVS | accuracy: 97.50% -> 97.92% | Helped |
| increase SNN timesteps from 2 to 16 (ablation) | SpikePoint · Daily DVS | accuracy: 92.75 -> 97.92 | Helped |
| increase SNN timesteps from 2 to 16 (ablation) | SpikePoint · DVS ACTION | accuracy: 60.93 -> 90.6 | Helped |
| replace IF neurons with LIF neurons (ablation) | SpikePoint · DVS128 Gesture | accuracy: 97.78% -> 98.74% | Helped |
| use the proposed ResF residual block instead of no residual connection (ablation) | SpikePoint · DVS ACTION | accuracy | Helped |
| use the SpikePoint singular-stage architecture with both local and global feature extraction (ablation) | SpikePoint · DVS ACTION | accuracy | Helped |
| use a 0.5 s sliding window instead of the selected 1.5 s window (ablation) | SpikePoint · Daily DVS | accuracy: 80% | Hurt |
| rescale relative coordinates by standard deviation before rate coding (ablation) | SpikePoint · Daily DVS | mean relative error: 1.07 -> 0.26 | Helped |
Surprising findings
- Absolute-value encoding of relative coordinates outperformed [0,1] normalization on Daily DVS (97.63-97.92% vs 96.25-96.53%).
- The authors report that for DVS Action, the sampling start point at half the timestamps was used as label-free data construction, and preprocessing denoising was necessary for good accuracy.
Limitations
- Performance on HMDB51-DVS is limited by background noise in the event data, and random sampling picks up noise.
- DVS Action required custom preprocessing (denoising, using second half of stream) that may limit comparability.
- Accuracy on UCF101-DVS and HMDB51-DVS remains well below the best ANN results.
Experimental setup
Datasets: DVS128 Gesture Daily DVS DVS Action HMDB51-DVS UCF101-DVS
Benchmarks: DVS128 Gesture Daily DVS DVS Action HMDB51-DVS UCF101-DVS
Metrics: accuracy parameters (M) dynamic energy (mJ) static energy (mJ)
Baselines: TBR+I3D PointNet++ EV-VGCNN VMV-GCN SEW-ResNet Spikformer Spikingformer tdBN PLIF ST-EVNet I3D TANet
Builds on: PointNet++ ParametricLIF spikingjelly
Closest papers by meaning
- Leveraging Asynchronous Spiking Neural Networks for Ultra Efficient Event-Based Visual Processing (AAAI 2025 · similarity 0.90)
- CREST: An Efficient Conjointly-trained Spike-driven Framework for Event-based Object Detection Exploiting Spatiotemporal Dynamics (AAAI 2025 · similarity 0.89)
- Spiking Neural Network as Adaptive Event Stream Slicer (NeurIPS 2024 · similarity 0.89)
- SFOD: Spiking Fusion Object Detector (CVPR 2024 · similarity 0.89)
- Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · similarity 0.88)
- Fully Spiking Neural Networks for Unified Frame-Event Object Tracking (NeurIPS 2025 · similarity 0.88)
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.