3D cascade RCNN for object detection on point cloud

Last update: Dec 02, 2022

Overview

3D Cascade RCNN

This is the implementation of 3D Cascade RCNN: High Quality Object Detection in Point Clouds.

We designed a 3D object detection model on point clouds by:

Presenting a simple yet effective 3D cascade architecture
Analyzing the sparsity of the point clouds and using point completeness score to re-weighting training samples. Following is detection results on Waymo Open Dataset.

Results on KITTI

	Easy Car	Moderate Car	Hard Car
AP 11	90.05	86.02	79.27
AP 40	93.20	86.19	83.48

Results on Waymo

	Overall Vehicle	0-30m Vehicle	30-50m Vehicle	50m-Inf Vehicle
LEVEL_1 mAP	76.27	92.66	74.99	54.49
LEVEL_2 mAP	67.12	91.95	68.96	41.82

Installation

Requirements. The code is tested on the following environment:

Ubuntu 16.04 with 4 V100 GPUs
Python 3.7
Pytorch 1.7
CUDA 10.1
spconv 1.2.1

Build extensions

python setup.py develop

Getting Started

Prepare for the data.

Please download the official KITTI dataset and generate data infos by following command:

python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos tools/cfgs/kitti_dataset.yaml

The folder should be like:

data
├── kitti
│   │── ImageSets
│   │── training
│   │   ├──calib & velodyne & label_2 & image_2
│   │── testing
│   │   ├──calib & velodyne & image_2
|   |── kitti_dbinfos_train.pkl
|   |── kitti_infos_train.pkl
|   |── kitti_infos_val.pkl

Training and evaluation.

The configuration file is in tools/cfgs/3d_cascade_rcnn.yaml, and the training scripts is in tools/scripts.

cd tools
sh scripts/3d-cascade-rcnn.sh

Test a pre-trained model

The pre-trained KITTI model is at: model. Run with:

cd tools
sh scripts/3d-cascade-rcnn_test.sh

The evaluation results should be like:

2021-08-10 14:06:14,608   INFO  Car [email protected], 0.70, 0.70:
bbox AP:97.9644, 90.1199, 89.7076
bev  AP:90.6405, 89.0829, 88.4391
3d   AP:90.0468, 86.0168, 79.2661
aos  AP:97.91, 90.00, 89.48
Car [email protected], 0.70, 0.70:
bbox AP:99.1663, 95.8055, 93.3149
bev  AP:96.3107, 92.4128, 89.9473
3d   AP:93.1961, 86.1857, 83.4783
aos  AP:99.13, 95.65, 93.03
Car [email protected], 0.50, 0.50:
bbox AP:97.9644, 90.1199, 89.7076
bev  AP:98.0539, 97.1877, 89.7716
3d   AP:97.9921, 90.1001, 89.7393
aos  AP:97.91, 90.00, 89.48
Car [email protected], 0.50, 0.50:
bbox AP:99.1663, 95.8055, 93.3149
bev  AP:99.1943, 97.8180, 95.5420
3d   AP:99.1717, 95.8046, 95.4500
aos  AP:99.13, 95.65, 93.03

Acknowledge

The code is built on OpenPCDet and Voxel R-CNN.

3D cascade RCNN for object detection on point cloud

Related tags

Overview

3D Cascade RCNN

Results on KITTI

Results on Waymo

Installation

Getting Started

Prepare for the data.

Training and evaluation.

Test a pre-trained model

Acknowledge

Owner

Qi Cai

Official implementation of the NRNS paper: No RL, No Simulation: Learning to Navigate without Navigating

Open source Python module for computer vision

Official repository for Automated Learning Rate Scheduler for Large-Batch Training (8th ICML Workshop on AutoML)

PyTorch implementation of Advantage Actor Critic (A2C), Proximal Policy Optimization (PPO), Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation (ACKTR) and Generative Adversarial Imitation Learning (GAIL).

Code for 'Blockwise Sequential Model Learning for Partially Observable Reinforcement Learning' (AAAI 2022)

Code repository for "Free View Synthesis", ECCV 2020.

Internship Assessment Task for BaggageAI.

RSNA Intracranial Hemorrhage Detection with python

TSP: Temporally-Sensitive Pretraining of Video Encoders for Localization Tasks

[ICCV 2021] Official Pytorch implementation for Discriminative Region-based Multi-Label Zero-Shot Learning SOTA results on NUS-WIDE and OpenImages

Selective Wavelet Attention Learning for Single Image Deraining

Bayesian Meta-Learning Through Variational Gaussian Processes

Pytorch Implementation of PointNet and PointNet++++

ContourletNet: A Generalized Rain Removal Architecture Using Multi-Direction Hierarchical Representation

PyTorch implementation of DARDet: A Dense Anchor-free Rotated Object Detector in Aerial Images

[NeurIPS 2021] Garment4D: Garment Reconstruction from Point Cloud Sequences

SAAVN - Sound Adversarial Audio-Visual Navigation,ICLR2022 (In PyTorch)

Tensorflow 2 implementation of the paper: Learning and Evaluating Representations for Deep One-class Classification published at ICLR 2021

Self-supervised learning algorithms provide a way to train Deep Neural Networks in an unsupervised way using contrastive losses

The official implementation of A Unified Game-Theoretic Interpretation of Adversarial Robustness.