A Low-Shot Object Counting Network With Iterative Prototype Adaptation (ICCV 2023)

The official PyTorch implementation of the ICCV 2023 paper LOCA.

@InProceedings{Dukic_2023_ICCV,
    author    = {{\DJ}uki\'c, Nikola and Luke\v{z}i\v{c}, Alan and Zavrtanik, Vitjan and Kristan, Matej},
    title     = {A Low-Shot Object Counting Network With Iterative Prototype Adaptation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {18872-18881}
}

Setup

Install the libraries

To run the code, install the following libraries: PyTorch 1.11.0, Torchvision 0.12.0, scipy, numpy and PIL.

Download the dataset

Download the FSC147 dataset as instructed in its official repository. Make sure to also download the annotation_FSC147_384.json and Train_Test_Val_FSC_147.json and place them alongside the image directory (images_384_VarV2) in the directory of your choice.

Download the pretrained models

Download the few-shot and/or zero-shot pretrained models and place them in a directory of your choice.

Generate density maps (optional)

If you wish to train LOCA on the FSC147 dataset, you need to run the density map generation script as follows:

python utils/data.py --data_path <path_to_your_data_directory> --image_size 512

where <path_to_your_data_directory> is the path to the dataset created in the previous step.

Training

The training code is located in the train.py script. It is adapted for distributed training on multiple GPUs. An example of how to run training on a SLURM-based system is shown in train_few_shot.sh and train_zero_shot.sh. Make sure to modify the paths in --data_path and --model_path to point to your data and model directories. To run locally on multiple GPUs, use torchrun (e.g., torchrun main.py --nproc_per_node=2 train.py <LOCA args (see SLURM example)>).

Evaluation

To evaluate LOCA, use the evaluate.py script (used the same way as train.py). You might want to run the evaluation on a single GPU. You can do that for the provided pretrained models as follows:

few-shot:

CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node=1 evaluate.py --model_name loca_few_shot --backbone resnet50 --swav_backbone --reduction 8 --image_size 512 --num_enc_layers 3 --num_ope_iterative_steps 3 --emb_dim 256 --num_heads 8 --kernel_dim 3 --num_objects 3 --pre_norm

zero-shot:

CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node=1 evaluate.py --model_name loca_zero_shot --backbone resnet50 --swav_backbone --reduction 8 --image_size 512 --num_enc_layers 3 --num_ope_iterative_steps 3 --emb_dim 256 --num_heads 8 --kernel_dim 3 --num_objects 3 --pre_norm --zero_shot

djukicn/loca