The most used pre-trained networks for computer vision (e.g. ResNet50) are trained on 3 channels (RGB). At the same time, many cameras used in robotics return RGB-D outputs, that is including depth information.

Fine tuning 3-channels pre-trained network on RGB-D outputs force us to give up one of the channels. Instead, is there a publicly available vision pre-trained network trained on RGB-D (4) channels?

  • $\begingroup$ As it is formulated now, this question may be a better fit at Open Data Stack Exchange. Nevertheless, instead of looking for RGB-D pretrained models, you could also add an extra channel dimension to the convolution filters of a pretrained RGB model and initialize randomly only that dimension before fine-tuning on your data. This would allow you to profit from a pretrained model while not giving up the extra channel. $\endgroup$ – noe Dec 21 '20 at 21:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.