2nd Place for Kaggle_PKU_Baidu Solution

Below you can find a outline of how to reproduce our solution for the Peking University/Baidu - Autonomous Driving competition.

The solution is detailed in Solution.

ARCHIVE CONTENTS

One trained model can be found at Google Drive. It achieves 0.112/0.118 on (private/public Leaderboard).

Installation

Requirements

We have tested the following versions of OS and softwares:

• OS: Ubuntu 16.04 LTS

• nvidia drivers v.384.130

• 4 x NVIDIA GeForce GTX 1080

• Python 3.6.9

• CUDA 9.0

• cuddn 7.4

• pytorch 1.1 (or +)

• GCC(G++): 4.9/5.3/5.4/7.3

• mmdet: 1.0.rc0+d3ca926
  (Or you can install the mmdet from the uploaded files. The newest mmdet 1.4+ has different API in calling mmcv. Hence, we would recommend install the mmdet from the uploaded files using: python setup.py install)

Configurations

All the data and network configurations are in the .config file: ./configs/htc/htc_hrnetv2p_w48_20e_kaggle_pku_no_semantic_translation.py

Dataset setup

Download the dataset from Kaggle and modify the config file to where you download the data data_root = '/data/Kaggle/pku-autonomous-driving/'

The newly added dataloader from mmdet/datasets/kaggle_pku.py will generate a corresponding annotation json file.

Running the code

training

In the tools folder, we provide the scripts for training:

• single gpu training: python train_kaggle_pku.py : currently single gpu training does not support validation evaluation. It is recommended to use the below

• multi-gpu training: CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 python -m torch.distributed.launch --nproc_per_node=6 train_kaggle_pku.py --launcher pytorch

• validation is only written for distributed training. If you only have a single gpu, you can do something like: CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 train_kaggle_pku.py --launcher pytorch

It is also recommended to load some pretrained model from mmdet.

Note: kaggle_apollo_combined_6691_origin.json is the annotation file from the combination of ApolloScape and Kaggle (We also cleaned up the noisy images with mesh overlay visualisation). Alternatively, you can use train.json from the Kaggle file. We have uploaded a json file kaggle_apollo_combined_6691_origin.json to google drive as: （We have encoded an absolute path, so to use the clean up data, you need to change the filepath according to your data location--> both for Kaggle and Apolloscape)

https://drive.google.com/open?id=1gEK7aGvTSAi8o2Jq3PgFky_YYX-HpJDt

Inference

• single model inference: the inference time is around 3 fps. python test_kaggle_pku.py and it will generate a single model .csv file. post processing

Model merging

python tools/model_merge.py in this script, you need to set the corresponding generated predicted pickle file.

Cite

If you find this repo helpful, we would appreciate if you cite the following paper.

@article{NMR6D2020,
  title   = {Neural Mesh Refiner for 6-DoF Pose Estimation},
  author  = {Di Wu and Yihao Chen and Xianbiao Qi and Yongjian Yu and Weixuan Chen and Rong Xiao},
  journal= {arXiv preprint 	arXiv:2003.07561},
  year={2020}
}

@InProceedings{Wu_2019_CVPR_Workshops,
author = {Wu, Di and Zhuang, Zhaoyong and Xiang, Canqun and Zou, Wenbin and Li, Xia},
title = {6D-VNet: End-to-End 6-DoF Vehicle Pose Estimation From Monocular RGB Images},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}

Contact

This repo is currently maintained by Di Wu (@stevenwudi) and Yihao Chen (@cyh1112).