This repository is the non-official pytorch implementation of a WACV 2021 Paper "Class-agnostic Few-shot-Object-Counting". Link

In Proc. IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021 Shuo-Diao Yang, Hung-Ting Su, Winston H. Hsu, Wen-Chin Chen^*

Our code has been implemented on Python 3.8 and PyTorch 1.8.1+cu101. Please follow the instructions to setup your environment. See other required packages in requirements.txt.

conda create --name CFOCNet python=3.8
conda activate CFOCNet
pip install -r requirements.txt
pip install "git+https://github.com/philferriere/cocoapi.git#egg=pycocotools&subdirectory=PythonAPI"

If you have problem about installing cocoapi, come here to find the official documentation.

• CFOCNet_demo.ipynb This notebook tests the detail implementations of CFOCNet, giving insights such as how the size of each tensor changes across each stage.
• model This directory contains all related modules of our CFOCNet implementation
• Eval_Result This directory contains the ideal results during evaluation stage, where an example's predicted count and the density map aligns with the groundtruth.

We train and evaluate our methods on COCO dataset 2017.
Please follow the instruction here to download the COCO dataset 2017
structure used in our code will be like :

$PATH_TO_DATASET/
├──── images
│    ├──── train2017
│             |──── 118287 images (.jpg)
│
│    ├──── test2017
│             |──── 40670 images (.jpg)
│
│    ├──── val2017
│             |──── 5000 images (.jpg)
│
├──── annotations
│    ├──── captions_train2017.json
│
│    ├──── captions_val2017.json
│
│    ├──── instances_train2017.json
|
│    ├──── instances_val2017.json
│
│    ├──── person_keypoints_train2017.json
│
│    ├──── person_keypoints_val2017.json

After downloading the data, please navigate to our repository.
Then, modify the variable "coco_path" in line 8 in crop.py to your COCO dataset path.

cd CODE_DIRECTORY
python data/crop.py

After performing the above instructions, the structure of your coco dataset will be like :

$PATH_TO_DATASET/
├──── images
│    ├──── train2017
│             |──── 118287 images (.jpg)
│
│    ├──── test2017
│             |──── 40670 images (.jpg)
│
│    ├──── val2017
│             |──── 5000 images (.jpg)
│
│    ├──── crop
│             |──── 80 directories which store all categories 500 images in coco dataset 2017 (for references images)
│
├──── annotations
│    ├──── captions_train2017.json
│
│    ├──── captions_val2017.json
│
│    ├──── crop.json
│
│    ├──── instances_train2017.json
|
│    ├──── instances_val2017.json
│
│    ├──── person_keypoints_train2017.json
│
│    ├──── person_keypoints_val2017.json

• Please go to config.yaml to change the configs under "train".
  
• To setup the training process, model configurations such as epochs, batch_size, and result_path can be tuned, all of which are stored in config.yaml
  
• Modify run.sh to setup the training process python main.py --config=config.yaml --doc=doc_name --train.
• doc_name can be any string you want.
• Execute the bash script through the below command:

cd CODE_DIRECTORY
bash run.sh

• After running the code, you will find your training logs under CODE_DIRECTORY/exp/logs/doc_name

• Please go to config.yaml to change the configs under "eval".
  
• To setup the testing process, the configurations such as checkpoint, sample, and image_folder can be tuned, all of which are stored in config.yaml.
  
• Modify run.sh to setup the testing process python main.py --config=config.yaml --doc=doc_name --test
• doc_name can be any string you want.
• Execute the bash script through the below command

cd CODE_DIRECTORY
bash run.sh

• After running the code, you will find your testing logs under "CODE_DIRECTORY/exp/logs/doc_name".

• The runner architecture is from NCSNv2.
• We crop 500 reference images for each categories.
• For query image, instead of random crop, we resize it with aspect ratio and padding to 256 x 256.
• The default setting in our code is 5-shot learning, where each query image has 5 reference images to learn.

• Great thanks to the contributive discussions on the reproduction details with
  the author of Class-Agnostic Few-Shot Object Counting, Shuo-Diao Yang,
  the author of Bilinear Matching Network, Min Shi,
  and the author of Learning to Count Anything: Reference-less Class-agnostic Counting with Weak Supervision, Michael Hobley.
  Their feedbacks have been really helpful, allowing us to replicate the model successfully.