Centernet: Many false detections from pascal modelzoo in webcam demo

[lih627]

When i run demo in webcam, and I use pretrained demo in model zoo, the results looks like that:

There are so many false detections . And with some other models, it detects nothing.

[lih627]

I don't know what caused this problem.
I use torch=0.4.1 torchvision=0.2.1

[lih627]

refer to #38

[lih627]

And this is the command

python demo.py ctdet --demo webacm --load_model ../models/ctdet_pascal_dla_384.pth

This situation only appears on the model pre-trained in PASCAL dataset..
By the way, how to use --arch

[xingyizhou]

The current implementation only support coco model in demo.py. To use pascal demo, you will need to change here. The reason why you got random predictions is that the last layer of the pascal model contains 20 channels and the created model with random initialization contains 80 channels. The pretrained weights are dropped.

[xingyizhou]

python demo.py ctdet --demo webacm --load_model ../models/ctdet_xxx --arch resdcn_18

[lih627]

Thanks a lot!@xingyizhou

[ashkan-abbasi66]

The current implementation only support coco model in demo.py. To use pascal demo, you will need to change here. The reason why you got random predictions is that the last layer of the pascal model contains 20 channels and the created model with random initialization contains 80 channels. The pretrained weights are dropped.

I have done the following steps:
1- Save opts.py as opts_voc.py
2- Save demo.py as demo_voc.py
3- In demo_voc.py, change from opts import opts to from opts_voc import opts
4- In init method of opts_voc.py, set the entry with ctdet key to the following:

  def init(self, args=''):
    default_dataset_info = {
        # mean and std for PASCAL VOC (Image Net) dataset are found at
        # https://www.learnopencv.com/pytorch-for-beginners-semantic-segmentation-using-torchvision/
      'ctdet': {'default_resolution': [512, 512], 'num_classes': 20, #80
                'mean': [0.485, 0.456, 0.406], 'std': [0.229, 0.224, 0.225],
                'dataset': 'pascal'},...# other things

After downloading a model that was trained over Pascal such as ctdet_pascal_dla_384, you can easily run demo_voc.py with the following command:
python demo_voc.py ctdet --demo ../images --load_model ../models/ctdet_pascal_dla_384.pth --dataset pascal