track2_aicity_2021's Introduction

A STRONG BASELINE FOR VEHICLE RE-IDENTIFICATION

This paper is accepted to the IEEE Conference on Computer Vision and Pattern Recognition Workshop(CVPRW) 2021

This repo is the official implementation for the paper A Strong Baseline For Vehicle Re-Identification in Track 2, 2021 AI CITY CHALLENGE.

I.INTRODUCTION

Our proposed method sheds light on three main factors that contribute most to the performance, including:

Minizing the gap between real and synthetic data
Network modification by stacking multi heads with attention mechanism to backbone
Adaptive loss weight adjustment.

Our method achieves 61.34% mAP on the private CityFlow testset without using external dataset or pseudo labeling, and outperforms all previous works at 87.1% mAP on the Veri benchmark.

II. INSTALLATION

pytorch>=1.2.0
yacs
apex (optional for FP16 training, if you don't have apex installed, please turn-off FP16 training by setting SOLVER.FP16=False)

$ git clone https://github.com/NVIDIA/apex
$ cd apex
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

python>=3.7
cv2

III. REPRODUCE THE RESULT ON AICITY 2020 CHALLENGE

Download the Imagenet pretrained checkpoint resnext101_ibn, resnet50_ibn, resnet152

1.Train

Prepare training data
- Convert the original synthetic images into more realistic one, using Unit repository
- Using Mask-RCNN (pre-train on COCO) to extract foreground (car) and background, then we swap the foreground and background between training images.
Vehicle ReID Train multiple models using 3 different backbones: ResNext101_ibn, Resnet50_ibn, Resnet152