SAFECount

October 10, 2023 · View on GitHub

Official PyTorch Implementation of Few-shot Object Counting with Similarity-Aware Feature Enhancement, Accepted by WACV 2023.

Image text

1. Quick Start

1.1 FSC147 in Original Setting

Create the FSC147 dataset directory. Download the FSC147 dataset from here. Unzip the file and move some to ./data/FSC147_384_V2/. The FSC147 dataset directory should be as follows.

|-- data
    |-- FSC147_384_V2
        |-- images_384_VarV2
        |-- gt_density_map_adaptive_384_VarV2
        |-- train.json
        |-- val.json
        |-- test.json
        |-- fold0.json
        |-- fold1.json
        |-- fold2.json
        |-- fold3.json

cd the experiment directory by running cd ./experiments/FSC147/.
Train, eval, or test by running:

(1) For slurm group: sh train.sh #NUM_GPUS #PARTITION, sh eval.sh #NUM_GPUS #PARTITION, or sh test.sh #NUM_GPUS #PARTITION.

(2) For torch.distributed.launch: sh train_torch.sh #NUM_GPUS #GPU_IDS, sh eval_torch.sh #NUM_GPUS #GPU_IDS, or sh test_torch.sh #NUM_GPUS #GPU_IDS, e.g., train with GPU 1,3,4,6 (4 GPU in total): sh train_torch.sh 4 1,3,4,6.

Note: During eval or test, please set config.saver.load_path to load the checkpoints.
Results and checkpoints.

Platform	Val MAE	Val RMSE	Test MAE	Test RMSE	Checkpoints
8 GPUs (NVIDIA Tesla V100 16GB)	14.42	51.72	13.56	91.30	here

1.2 FSC147 in Cross Validation Setting

Here we provide one example (3 shot, fold0 as val set), and others are similar.

Create the FSC147 dataset directory, the same as in 1.1.
cd the experiment directory by running cd experiments/FSC147_fold/3shot/fold0/.
Train or eval by running:

(1) For slurm group: sh train.sh #NUM_GPUS #PARTITION or sh eval.sh #NUM_GPUS #PARTITION.

(2) For torch.distributed.launch: sh train_torch.sh #NUM_GPUS #GPU_IDS or sh eval_torch.sh #NUM_GPUS #GPU_IDS.

Note: During eval or test, please set config.saver.load_path to load the checkpoints.
Results. Training on 8 GPUs (NVIDIA Tesla V100 16GB) results in following performance.

Shot	Val Fold	Val MAE	Val RMSE	Shot	Val Fold	Val MAE	Val RMSE
1	0	15.62	51.20	3	0	13.84	43.99
1	1	7.14	15.77	3	1	6.29	13.89
1	2	14.05	92.48	3	2	11.19	86.81
1	3	13.88	38.38	3	3	11.73	33.85

1.3 Cross-dataset Generalization (FSC147 to CARPK)

Create the FSC147 & CARPK dataset directory, the same as in 1.1 & 1.4.
cd the experiment directory by running cd experiments/FSC147_to_CARPK/.
Pretrain, finetune, or eval. The pretrain, finetune, or eval are similar to FSC147.
Results. Training on 8 GPUs (NVIDIA Tesla V100 16GB) results in following performance.

MAE (pretrain on FSC147)	RMSE (pretrain on FSC147)	MAE (finetune on CARPK)	RMSE (finetune on CARPK)
17.78	20.95	4.91	6.32

The train and eval of class-specific counting are similar to FSC147. Here we only provide the construction of the dataset directory. The checkpoints (trained on 8 NVIDIA Tesla V100 16GB GPUs) and the corresponding results are given.

	CARPK	PUCPR+	UCSD	Mall	ShanghaiTech PartA	ShanghaiTech PartB
MAE	4.91	2.24	1.01	1.77	74.36	9.75
RMSE	6.32	3.44	1.34	2.24	121.15	15.87
checkpoints	here	here	here	here	here	here

1) CARPK

Download the CARPK dataset from here. Unzip the file and move some to ./data/CARPK_devkit/. The CARPK dataset directory should be as follows.

|-- data
    |-- CARPK_devkit
        |-- Images
        |-- gen_gt_density.py
        |-- train.json
        |-- test.json
        |-- exemplar.json

run python gen_gt_density.py to generate ground-truth density map. The ground-truth density map will be saved to ./data/CARPK_devkit/gt_density_map/.

2) PUCPR+

Download the PUCPR+ dataset from here. Unzip the file and move some to ./data/PUCPR+_devkit/. The PUCPR+ dataset directory should be as follows.

|-- data
    |-- PUCPR+_devkit
        |-- Images
        |-- gen_gt_density.py
        |-- train.json
        |-- test.json
        |-- exemplar.json

run python gen_gt_density.py to generate ground-truth density map. The ground-truth density map will be saved to ./data/PUCPR+_devkit/gt_density_map/.

3) UCSD

Download the UCSD dataset from here. Unzip the file and move some to ./data/UCSD/. The UCSD dataset directory should be as follows.

|-- data
    |-- UCSD
        |-- ucsdpeds_vidf
        |-- gen_gt_density.py
        |-- train.json
        |-- test.json
        |-- exemplar.json
        |-- mask.png

We use the annotations in vidf-cvpr.zip, which corresponds to 10 directories (from ucsdpeds_vidf/video/vidf/vidf1_33_000.y/ to ucsdpeds_vidf/video/vidf/vidf1_33_009.y/). Merge all images under these 10 directories to ucsdpeds_vidf/video/vidf/. Other directories could be removed.
run python gen_gt_density.py to generate ground-truth density map. The ground-truth density map will be saved to ./data/UCSD/gt_density_map/.

4) Mall

Download the Mall dataset from here. Unzip the file and move some to ./data/Mall/. The Mall dataset directory should be as follows.

|-- data
    |-- Mall
        |-- frames
        |-- gen_gt_density.py
        |-- train.json
        |-- test.json
        |-- exemplar.json
        |-- mask.png

run python gen_gt_density.py to generate ground-truth density map. The ground-truth density map will be saved to ./data/Mall/gt_density_map/.

5) ShanghaiTech

Download the ShanghaiTech dataset from here. Unzip the file and move some to ./data/ShanghaiTech/. The ShanghaiTech dataset directory should be as follows.

|-- data
    |-- ShanghaiTech
        |-- part_A
            |-- train_data
                |-- images
            |-- test_data
                |-- images
            |-- gen_gt_density.py
            |-- train.json
            |-- test.json
            |-- exemplar.json
        |-- part_B
            |-- train_data
                |-- images
            |-- test_data
                |-- images
            |-- gen_gt_density.py
            |-- train.json
            |-- test.json
            |-- exemplar.json

run python gen_gt_density.py to generate ground-truth density map. Note that you should run this twice for both part_A and part_B. The ground-truth density map will be saved to (1) ./data/ShanghaiTech/part_A/train_data/gt_density_map/, (2) ./data/ShanghaiTech/part_A/test_data/gt_density_map/, (3) ./data/ShanghaiTech/part_B/train_data/gt_density_map/, (4) ./data/ShanghaiTech/part_B/test_data/gt_density_map/.

2. Learn More About Our Methods

We provide two datasets: custom_dataset.py & custom_exemplar_dataset.py, and three models: safecount.py, safecount_exemplar.py, & safecount_crossdataset.py. They should be cooperated and used as follows.

dataset	model	circumstance
custom_dataset.py	safecount.py	The support images are parts of the query image, and annotated by bounding boxes, e.g., FSC147.
custom_exemplar_dataset.py	safecount_exemplar.py	The support images are sampled then fixed, and not parts of the query image, e.g., CARPK, PUCPR+, UCSD, Mall, and ShanghaiTech.
custom_dataset.py	safecount_crossdataset.py	In cross-dataset generalization, pretraining a model on FSC147 and finetuning on class-specific counting.

3. Questions

CUDA Out of Memory.

(1). Choose a smaller image size (config.dataset.input_size).

(2). Set a smaller exemplar number (config.dataset.shot for FSC147, config.dataset.exemplar.num_exemplar for class-specific counting).

(3). Set a larger out_stride (config.net.kwargs.backbone.out_stride), but you also need to revise the Regressor (in models.utils.py) to upsample the feature to the original image size.