Awesome

UniFS

Introduction

This repo contains the official PyTorch implementation of our ECCV'2024 paper UniFS: Universal Few-shot Instance Perception with Point Representations.

Data Preparation

We evaluate our models on COCO-UniFS benchmark. This benchmark is built upon several existing datasets, including MSCOCO and MISC.

The COCO-UniFS benchmark provides dense annotations for four fundamental few-shot computer vision tasks: object detection, instance segmentation, pose estimation, and object counting. The annotations for object detection and instance segmentation are directly taken from the MSCOCO dataset, which provides bounding box and per-instance segmentation mask annotations for 80 object categories. For pose estimation, we extend the MSCOCO dataset by adding instance-level keypoint annotations for 34 object categories from the MISC dataset. The MISC dataset was originally designed for multi-instance semantic correspondence, and we adapted it to fit the few-shot pose estimation task. The dataset split follows DeFRCN.

• Unzip the downloaded COCO-UniFS data-source to datasets and put it into your project directory:

    ...
    datasets
      | -- coco (trainval2014/*.jpg, val2014/*.jpg, annotations/*.json)
      | -- unifs_cocosplit
    unifS
    tools
    ...
  

Training and Evaluation

• To reproduce the FSIS/gFSIS results on COCO

  bash run/run_unify.sh r101 dcfs 8
  

Visualization Results

The baseline DeFRCN may tend to incorrectly recognize positive object as background (middle two rows) due to the biased classification. This problem is greatly alleviated using our proposed method (DCFS).

Acknowledgement

This repo is developed based on DCFS, DeFRCN and Detectron2.

License

UniFS is freely available for free non-commercial use, and may be redistributed under these conditions. For commercial queries, please contact Mr. Sheng Jin (jinsheng13[at]foxmail[dot]com). We will send the detail agreement to you.