JREion/Prompt_Tuning_Datasets_with_Foreground

⭐ Dataset Introduction

The standard datasets (except ImageNet) used for CLIP-based Prompt Tuning research (e.g., CoOp).

Based on the original datasets, this repository adds foreground segmentation masks (generated by SEEM) of all raw images.

• For the foreground masks, the RGB value of the foreground region is [255, 255, 255], and the background region is [0, 0, 0].

• The shorter side is always fixed to 512 px, and the scaling ratio is the same as that of the corresponding raw image.

We provide masks for the following datasets: ImageNet, Caltech101, Oxford Pets, StanfordCars, Flowers102, Food101, FGVC Aircraft, SUN397, DTD, EuroSAT and UCF101.

If you only want to download the mask data (do not contain raw images), please download from [_OPTIONAL_FOREGROUND_MASK_ONLY_DATA] folder.

🏷️ Scope of Application

Datasets are suitable for training and improving foreground-supervised prompt tuning methods. For example:

• FVG-PT: Adaptive Foreground View-Guided Prompt Tuning for Vision-Language Models   [GitHub]   [Paper]

Also, they are fully compatible with other original prompt tuning approaches.

⚙ Data Preparation

The datasets include the original images, the split_zhou_xxx.json annotations, and foreground masks.

The mask directory is located under the dataset root, and its internal subpath is consistent with the image directory, e.g.:

• Image directory: ./flowers-102/oxford_flowers/jpg
• Mask directory: ./flowers-102/mask/oxford_flowers/jpg

Additionally, you can prepare ImageNet dataset from: [Raw Images] [annoations] [val conversion script]

NOTE: You can build the file tree by referring to the [FVG-PT repository].

🖥︎ When You Write Your Own Code ...

1. You can refer to ./Dassl.pytorch in the FVG-PT repository to build a DataLoader that can pass masks.

   mask_path = join(dataset_root, "mask", image_path_suffix)
   

2. During pre-processing, the mask input needs to be resized, e.g.:

   def _transform_pair(self, img, mask):
           if mask.size != img.size:
               mask = TF.resize(
                   mask,
                   [img.height, img.width],
                   interpolation=InterpolationMode.NEAREST
               )
   

Acknowledgements

Our repository is built based on DPC, FVG-PT, DAPT and zhengli97.