Architecture

NDConv

(a) shows the sampling locations of conventional deformable conv, and (b) demonstrates the sampling points of our Normed-Deformable conv.

Due to the extra constrain of NDLoss, the sampling points tend to be more likely in shape of ellipse.

Data Preparation

Download the datasets ShanghaiTech A, ShanghaiTech B, UCF-QNRF Then generate the density maps via generate_density_map_perfect_names.py.

change config.py to use the path of your dataset.

Pretrained model

bash download_models.sh

Train

sh run_QNRF_crop.sh

For CSRNet⋆(Baseline) (ie, CSRNet with one traditional Deformable Layer), just set extra_loss=0 in run_QNRF_crop.sh, then the NDLoss will be discarded during training.

Training with Larger batchsize ?

Mention: Due to the limitation of GPU (only a GPU with 8Gb RAM....), we set a very small batch size, and the image size is relatively small. If your GPU capacity is big enough, you can set larger batch size and use QNRF_large to train our model, the performance will be much better.

Test

Add opt.test_model_name=/path_of_eval_model in test.py and run

python test.py --net_name='csrnet_deform_var' --gpu_ids='0'