IC-FPS: Instance-Centroid Faster Point Sampling Module for 3D Point-base Object Detection
News
- 2023.4.6 create README.md
Main Result
Waymo detection validation
Quantitative comparison experiments on the Waymo val set for 3D object detection. 20% of training set ( 32K) is used for
training.
Quantitative comparsion experiments on Waymo val set for 3D object detection. † indicates that the whole scene is divided into
four parts to accelerate SA layer. Bold texts are our results and the best results are underlined.
ONCE detection validation
Quantitative comparison experiments on ONCE validation set. Bold texts are our results, underlined texts are the best results.
Performance indexes follow the same configurations given in IA-SSD
nuScenes detection validation
Visualization results
Visulization results of Waymo dataset Vehicle. (a)ground-truth, (b) IA-SSD, (c) IC-FPS+IA-SSD. Red boxes represent differences of various results.
Complementary visualization of Waymo Dataset Pedestrian, (a) ground truth, (b) IA-SSD, (c) IC-FPS + IA-SSD. Bounding boxes in red are the difference, green are the Pedestrian.
Complementary visualization of Waymo Dataset Cyclist. (a) ground truth, (b) IA-SSD, (c) IC-FPS + IA-SSD. Boundingboxes in red are the difference, purple are the cyclist.
Complementary visualization of our methods on nuScenes Dataset.
Use IC-FPS
install and date preparation
For environment installation method, please refer to IA-SSD.
beachmark Evaluation and Training
# training example for IC-FPS
cd tools && scripts/dist_train.sh 8 --cfg_file cfg/waymo_models/PTC.yaml --extra_tag IC-FPS
#evaluation example for IC-FPS
cd tools && scripts/dist_test.sh 8 --cfg_file cfg/waymo_models/PTC.yaml --ckpt YOUR_WEIGHT --extra_tag IC-FPSAcknowlegement
We sincerely thank the authors of IA-SSD, SASA, 3DSSD for open sourcing their methods.