search

Huangying-Zhan / DF-VO

Depth and Flow for Visual Odometry
https://df-vo.readthedocs.io
MIT License
556 stars 117 forks source link

Confusion about the post-processing in lite_flow_net.py #10

Closed DecaYale closed 2 years ago

DecaYale commented 4 years ago

Hi, I noticed that you add some post-processing in the forward() function of LiteFlowNet:

        for i in flows:
            flows[i] = flows[i] * (20.0 * (0.5 ** (i-1)))
            # _, _, out_h, out_w = flows[i].shape
            # flows[i] = torch.nn.functional.interpolate(input=flows[i], size=(raw_h, raw_w), mode='bilinear', align_corners=False)
            # flows[i][:, 0, :, :] *= float(raw_w) / float(out_w)
            # flows[i][:, 1, :, :] *= float(raw_h) / float(out_h)

Why is this necessary? 20 seems to be a strange number and I am confused a little bit .

Huangying-Zhan commented 4 years ago

sorry for late reply, it is taken from this re-implementation. https://github.com/sniklaus/pytorch-liteflownet