Reproduce “Optimisation of the PointPillars network for 3D object detection in point clouds” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Joanna Stanisz, Konrad Lis, Tomasz Kryjak, Marek Gorgon
The PointPillars network was optimized for 3D object detection in point clouds, achieving a 16-fold reduction in model size and a maximum 10% drop in detection accuracy through combined quantisation and pruning techniques.
Key insights
- Aggressive quantisation, especially to 2-bit integers for the network's Backbone, enables substantial model size reduction crucial for embedded systems like FPGAs.
- Re-training pruned networks with unfrozen weights effectively mitigates the performance degradation caused by high sparsity levels, even at 80% pruning.
- Quantising activation functions to lower precision (e.g., INT4 for ReLU) can sometimes yield comparable or better results than higher precisions, highlighting non-linear effects of activation quantisation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.