Calibrating_count

The code for ICCV2023 paper: Calibrating Uncertainty for Semi-Supervised Crowd Counting

Semi-supervised crowd counting is an important yet challenging task. A popular approach is to iteratively generate pseudo-labels for unlabeled data and add them to the training set. The key is to use uncertainty to select reliable pseudo-labels. In this paper, we propose a novel method to calibrate model uncertainty for crowd counting. Our method takes a supervised uncertainty estimation strategy to train the model through a surrogate function. This ensures the uncertainty is well controlled throughout the training. We propose a matching-based patch-wise surrogate function to better approximate uncertainty for crowd counting tasks. The proposed method pays a sufficient amount of attention to details, while maintaining a proper granularity. Altogether our method is able to generate reliable uncertainty estimation, high quality pseudolabels, and achieve state-of-the-art performance in semisupervised crowd counting.

1. Environment setup

Using conda:

conda env create -f crowd_test.yml

2. Preparing dataset for ShanghaiTech Part-A

Download the ShanghaiTech Part-A dataset.

makedir part_A

move data into folder part_A

.
├── ...
├── part_A                   
│   ├── train_data          
│   ├── test_data        
│   └──               
└── ...

makedir part_A/uncertain_data

makedir part_A/uncertain_label

makedir part_A/uncertain_test

makedir part_A/uncertain_data_5

makedir part_A/uncertain_data_10

makedir part_A/uncertain_data_40

python patch_gen_A.py # crop whole images into little patches.

python patch_move.py # move cropped patches into corresponding folders

2. Training and evaluating on 10% labeled ShanghaiTech Part-A

python train_semi_A.py --data_root ../CrowdCounting-P2PNet/part_A/  \
   --dataset_file SHHA_partial  \
   --epochs 1000  \
   --lr_drop 3500   \
  --output_dir ./res/SHHA/logs  \
   --checkpoints_dir ./res/SHHA/weights   \
  --tensorboard_dir ./res/SHHA/logs   \
  --lr 0.0001   \
  --lr_backbone 0.00001  \
   --batch_size 8   \
  --eval_freq 1  \
   --gpu_id 1  \
   --un_weight 0.3  \
   --end_pro 0.4  \
    --label_pro 10 

end_pro here is the confidence threshold (1-uncertainty) label_pro is the proportion of labeled data (10 for 10%. 5 for 5%, 40 for 40%)

The pretrained VGG weights can be downloaded from https://download.pytorch.org/models/vgg16_bn-6c64b313.pth