We proposed the ID-discriminative Embedding (IDE) in [Arxiv 2016] and [ECCV 2016], [Code] .
Later, we introduced the Part-based Convolutional Baseline (PCB) in [ECCV 2018], [Code] .
A tiny, friendly, strong baseline [Code] .
We studied the domain adaptive re-ID problem. A representatives work is CamStyle [CVPR 2018, TIP 2019], [Code] .

Generative Models for Discriminative Learning

For style-level domain adaptation, we designed SPGAN for person re-ID [CVPR 2018], [Code] . The GAN generated images are effective for discriminative learning.
We proposed DG-net to generate pedestrians with various appearances/structures to augment the real-world data in discriminative learning [CVPR 2019], [Code] .

Multi-object Tracking

We designed an efficient MOT system with near real-time performance, named JDE [Arxiv 2019], [Code] .
We also studied the inherent difference between re-ID and MTMCT, and proposed the locality aware appearance metric (LAAM) [Arxiv 2019], [Code] .

Data Augmentation

We introduced Random Erasing for data augmentation [AAAI 2020], [Code] .
Random Erasing is effective in image classification such as CIFAR and ImageNet, and re-ID.
There are some important third-party implementations, such as Official Torchvision and Python Augmentor.

Domain Adaptive Semantic Segmentation

We released the code for CLAN, a category-level alignment method [CVPR 2019], [Code] .
We also released our code for MMAN that uses the adversarial loss [ECCV 2018], [Code] .