Closed TsainGra closed 4 years ago
Yes, the keypoints is trained along with entire system for specific dataset. So in case of moving gif keypoint detector is specialized on 4 leged cartonish aninals. So it is not surprising that it not working on birds. Depends on how much data you have you can try to finetune the system on birds.
Hello, I am trying to transfer motion of one bird to another. However, due to lack of data, I am trying to infer from using moving-gif weights provided with the code. After adding a white background to all frames, I tried to reconstruct the video giving the first frame of the driving video as the source image. I observed that apart from the generator not able to generate convincing images, sometimes the key points were estimated incorrectly. Is there any sort of relation in which the keypoint detector depends on the object (mostly quadrupeds in moving-gif) the network was trained on.