On Using ConvNeXt within the IICS/IIDS Framework for Person Re-ID

Introduction

This project explores the integration of ConvNeXt, a CNN-based network inspired by vision transformers, into the Intra and Inter Camera Similarity (IICS) and Intra and Inter Domain Similarity (IIDS) frameworks for unsupervised person Re-ID. Building upon IICS/IIDS framework that generates pseudo labels through intra and inter stages and utilizing techniques such as Adaptive Instance and Batch Normalization (AIBN) and Transform Normalization (TNorm) to minimize intra-camera and inter-camera variations respectively, our work emphasizes the application of ConvNeXt as a feature extractor. ConvNeXt gets higher mAP and CMC on the Market1501 and MSMT17 datasets than most unsupervised learning methods. Furthermore, we explored the effect of AIBN and TNorm normalization techniques in ConvNeXt. We showed their effectiveness in reducing intra-camera and inter-camera variations if AIBN is inserted in the final stages (Stage 3 and stage 4) and TNorm layers are included after stage 1, stage 2, and stage 3. We also examined the effects of four ConvNeXt variants within the IICS/IIDS framework, emphasizing the advantages of using larger variants of ConvNeXt as a feature extractor for person Re-ID.

Presentation

Results

Below we include the results of our approach on three benchmarks. We compare the proposed method against several schemes that rely upon GANs, distribution alignment or psuedo-labeling. The proposed scheme achieves better results than existing approaches on both Market 1501 and MSMT17 datasets. IIDS (Xuan & Zhang, 2022) acheives better scores on DukeMTMC dataset. GANTs is a typo. This refers to GANs.

Results on Market 1501 dataset.

Results on DukeMTMC dataset.

Results on MSMT17 dataset.

Afterword

While we are interested in pushing the boundaries of technology, we are aware how these type of technologies and capabilities can be easily misused. Societal norms and legal structures are needed to ensure that such technologies are used for the good of all citizens. Please see https://eprints.ncrm.ac.uk/id/eprint/421/1/MethodsReviewPaperNCRM-011.pdf for a deeper discussion of ethical implications of this work. Local copy is provided here.

Code

Code for this project is available at https://github.com/vclab/convnext-person-reid.

Publication

For technical details please look at the following publication(s)