Deep-Learning-Push-Up-Counter VS DenseDepth

Hi ! I am currently working on a project which involves the recognition of tennis strokes in videos. I have already implemented the detection part using a pre-trained OpenPose model to find the body keypoints and an LSTM for the classification task. Now I am having some trouble counting those movements. I have seen various approaches, but I would like something lightweight, somehow related to this specific task. Here is a project which uses optical flows to detect patterns in the movements, but a tennis stroke is a more complex action and I don't know how to label the frames (swing, not swing maybe). Another solution would be to use the already computed keypoints to compute some angles, but I didn't find something related to tennis. Thank you!

For a long time, the state-of-the-art models for monocular depth estimation such as DORN and DenseDepth were built with convolutional neural networks. Recently, however, both transformer-based models such as DPT and GLPN, and diffusion-based models like Marigold have achieved remarkable results!

Instead of just doing grayscale for depth, you should consider using “monocular depth estimation” that actually tries to reconstruct a depth map from an RGB image. There are a variety of open-source libs available like this one.

After that I have to do the same for Python I think, and then I have to find out how to figure out how to use a library like https://github.com/ialhashim/DenseDepth or https://github.com/nianticlabs/monodepth2 for that GStreamer plugin (or element, still trying to grasp the terminology here)

Others model implemented like github.com/ialhashim/DenseDepth have constraints on input (I think only 4:3 is one of them, if I remember correctly)