MPD

Möbius Transform for Mitigating Perspective Distortions in Representation Learning (ECCV2024)

Chhipa, P.C., Chippa, M.S., De, K., Saini, R., Liwicki, M., Shah, M.: Möbius transform for mitigating perspective distortions in representation learning. European Conference on Computer Vision. (ECCV 2024)

Visit Project Website - https://prakashchhipa.github.io/projects/mpd/

1. Pretrained models - https://huggingface.co/prakashchhipa/MPD_SSL
2. ImageNet-PD benchmark dataset - https://huggingface.co/datasets/prakashchhipa/ImageNet-PD
3. Two minutes summary on MPD - https://prakashchhipa.github.io/projects/mpd/

Commands to Train Models

1. Supervised Training with ResNet50 and Mobius Probability 0.2

Train a ResNet50 model under a supervised setting with Mobius transformations applied with a probability of 0.2:

torchrun --nproc_per_node=2 train_supervised.py --model resnet50 --apply-mobius --forward-mobius --mobius-prob 0.2 --name supervised_mpd_rn50

2. SimCLR Pretraining with Mobius

Pretrain ResNet50 using the SimCLR framework with Mobius transformations:

python train_simclr.py --config configs/imagenet_train_epochs100_bs512.yaml

3. SimCLR Pretraining with Mobius and Padded Background

Pretrain ResNet50 with SimCLR, applying Mobius transformations and controlling the background with the mobius_background input:

python train_simclr.py --config configs/imagenet_train_epochs100_bs512.yaml

4. Finetune Downstream Task for SimCLR-Trained ResNet50

Fine-tune a ResNet50 model pretrained with SimCLR for a downstream task using Mobius transformations:

torchrun --nproc_per_node=5 train_downstream_task_from_ssl.py --model resnet50 --apply-mobius --apply-BGI --name downstream_simclr_rn50 --forward-mobius --mobius-prob 0.2 --ssl-checkpoint <checkpoint URL>

5. DINO Self-Supervised Pretraining

Pretrain a Vision Transformer (ViT) using DINO with Mobius transformations:

python -m torch.distributed.launch --nproc_per_node=8 main_dino_mobius.py --mobius_prob 0.8 --arch vit_small --data_path /path/to/imagenet/train --output_dir /path/to/saving_dir

6. DINO Self-Supervised Pretraining with Padded Background

Pretrain a Vision Transformer (ViT) using DINO with Mobius transformations and a padded background:

python -m torch.distributed.launch --nproc_per_node=8 main_dino_mobius_bgi.py --mobius_prob 0.8 --arch vit_small --data_path /path/to/imagenet/train --output_dir /path/to/saving_dir

In-progress Work

The source code for additional computer vision applications will be released later.

GitHub repository credits

simclr-pytorch pytorch-classfication