Density maps and people localization

[Zebrabeast]

Hello, your work looks really impressive, but I have some questions about the code: 1. In the paper, you mentioned that there would be density map generation later. From my understanding, the output in the test.py file should be a density map, but its size does not match the input's, so I am not sure if the output is indeed the density map referred to in the text. 2. Can your code identify the locations of people?

[LoraLinH]

Thanks for your interest. In the counting tasks, the output density map is typically smaller than the original image for two main reasons. First, the backbone is based on the VGG network, which inherently downsamples features by a factor of 16. Second, maintaining the density map at the original image size would result in an exponential increase in computational cost when calculating the loss. As a result, the model generates a smaller density map. For visualization purposes in the paper, I use cv2.resize to scale the model's output density map to the same size as the original image. Regarding your second question, our model currently does not support direct localization output. If needed, you could potentially apply a non-maximum suppression (NMS) algorithm to the output density map to extract localization points. However, I haven't tried this approach, and the results may be unstable.