Text-Guided Mixup Towards Long-Tailed Image Categorization

BMVC 2024

Richard Franklin, Jiawei Yao, Deyang Zhong, Qi Qian, Juhua Hu*

Model architecture with LFM used to extend the decision boundary of minor classes towards nearby classes

Requirements

• We recommend Linux for performance and compatibility reasons.
• 1 NVIDIA GPU for CIFAR10/100-LT, and 3 for Imagenet-LT and Places-LT. Each GPU that we trained the model with was an RTX 2080 Ti (11GB).
• Python dependencies are located in requirements.txt

Getting started

Datasets

• CIFAR100-LT
• CIFAR10-LT
• Places-LT
• Imagenet-LT

Training and evaluation

CIFAR100 dataset

python main.py --cfg config/general.yaml config/proposed/lfm-mms.yaml --gpu 0

CIFAR10 dataset

python main.py --cfg config/general.yaml config/cifar10.yaml config/proposed/lfm-mms.yaml --gpu 0

Places365 dataset

python main.py --cfg config/general.yaml config/places.yaml config/proposed/lfm-mms.yaml

Imagenet dataset

python main.py --cfg config/general.yaml config/imagenet.yaml config/proposed/lfm-mms.yaml

Acknowledgement

Yao and Hu's research is supported in part by NSF (IIS-2104270) and Advata Gift Funding. Zhong's research is supported in part by the Carwein-Andrews Graduate Fellowship and Advata Gift Funding. All opinions, findings, conclusions and recommendations in this paper are those of the author and do not necessarily reflect the views of the funding agencies.