Simple Question about object detection code

[AlexCo1d]

CoVGT/model/EncoderVid.py

Lines 56 to 71 in cbc9fa7

	def forward(self, video_o):
	bsize, numc, numf, numr, fdim = video_o.shape
	video_o = video_o.view(bsize, numc*numf, numr, fdim)
	roi_feat = video_o[:,:,:, :self.dim_feat]
	roi_bbox = video_o[:,:,:, self.dim_feat:(self.dim_feat+self.dim_bbox)]
	bbox_pos = self.bbox_conv(roi_bbox.permute(
	0, 3, 1, 2)).permute(0, 2, 3, 1)
	bbox_features = torch.cat([roi_feat, bbox_pos], dim=-1)
	bbox_feat = self.tohid(bbox_features)
	return bbox_feat

Hi, I am reading your code about object detection. I found the above one in your EncoderVid.py
Do you still remember why you choose 5 dimension (dim_bbox) for positional embedding? What is the source of this way? (Faster RCNN or Detectron)

Thank you for your prompt response! Thanks for your great work!

[doc-doc]

Hi, the fifth dimension denotes the relative bbox size: bbox_size/image_size(w*h), it is basically based on my previous relation grounding work :https://github.com/doc-doc/vRGV.