zuowanghanguang / pixor_pytorch

Introduction

This is an unofficial inplementation of [Bird's Eye View Object Detection Algorithm for self-driving Cars]PIXOR in Pytorch. A large part of this project is based on the work here. Thanks to @Ankita Kalra. This work is still ongoing.

Dependencies

• python3.5+
• Pytorch (tested on 0.4.1)
• opencv-python
• shapely
• matplotlib
• tensorboardX

Installation

1. Clone this repository.

Data Preparation

1. Download the 3D KITTI detection dataset from here. Data to download include:

   • Velodyne point clouds (29 GB): input data to VoxelNet
   • Training labels of object data set (5 MB): input label to VoxelNet
   • Camera calibration matrices of object data set (16 MB): for visualization of predictions
   • Left color images of object data set (12 GB): for visualization of predictions

2. Split the training set into training and validation set according to the protocol here. And rearrange the folders to have the following structure:

└── KITTI
       ├── training   <-- training data
       |   ├── image_2
       |   ├── label_2
       |   └── velodyne
       └── validation  <--- evaluation data
       |   ├── image_2
       |   ├── label_2
       |   └── velodyne
       |
       |__ train.txt
       |
       |__ val.txt
       |
       |__ trainval.txt

Train

$ python run_training.py

1. There is a pre-trained model for car in pretrained_models/model_90.pth.

inference

$ python run_inference.py

results

GT

prediction

TODO

• improve the performances
• reproduce results for Pedestrian and Cyclist
• provide SummaryWriter()
• provide run_evaluate.py