Track-Anything VS openpose

I was doing rotoscoping for a silhouette of a girl dancing in front of a building, then I saw this amazing tool: https://github.com/gaomingqi/Track-Anything

My question is, how can we modernize this pipeline? We've experimented a bit with the new SAM-based track-anything tool, and it seems promising, but we actually don't want to "track anything", we only want to track fishes. What would you do in 2023, to extract tracks of one specific class of object from long video datasets? I'm hoping for any advice at all.

The Track-Anything tool already implements this

Segment Anything for Video - Track Anything! 🤖 With this tool, you can automatically isolate objects, make edits using inpainting, and track objects with precision. It's a game-changer for creative projects. Even though it does not work well with the shadows yet, we expect a rapid evolution of these technologies. Github : https://github.com/gaomingqi/Track-Anything)

I would like to know if there are any guidelines to follow when adapting an existing Python project for my own use-case. Specifically, I want to customize the output of the Track-Anything project by incorporating my own processing steps. However, I do not want to import the entire codebase. Rather, I only want to import the minimum amount of code necessary to produce the same output with object tracking, without having to reimplement functions that are already available.

Here is the GitHub: https://github.com/gaomingqi/track-anything

Open Pose, a multi-person keypoint detection library for body, face, hands, and foot estimation [10], is used for posing generated characters;

OpenPose

Perhaps something like OpenPose for pose estimation?

OpenPose looks like what you're looking for, it seems to have plugins for Unity. I can't say I've used it though.

You are correct. Open Pose has two keypoints for the eyes and two more for the ears. By saying were the ears are you automatically influence the angle of the head. You can see more about it on this github page. Just scroll a tiny bit and you can see a gif of the nodes overlapped on humans

FROM nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04 # https://hub.docker.com/r/nvidia/cuda ENV DEBIAN_FRONTEND=noninteractive # install the dependencies for building OpenPose RUN apt-get update && # The rest is ignored for brevity. RUN pip3 install --no-cache-dir # The rest is ignored for brevity. # install cmake, clone OpenPose and download models RUN wget https://cmake.org/files/v3.20/cmake-3.20.2-linux-x86_64.tar.gz && \ # The rest is ignored for brevity. WORKDIR /openpose/build RUN alias python=python3 && cmake -DBUILD_PYTHON=OFF -DWITH_GTK=OFF -DUSE_CUDNN=ON .. # Build OpenPose. Cudnn 8 causes memory issues this is why we are using base with CUDA 10 and Cudnn 7 # Fix for CUDA 10.0 and Cudnn 7 based on the post below. # https://github.com/CMU-Perceptual-Computing-Lab/openpose/issues/1753#issuecomment-792431838 RUN sed -ie 's/set(AMPERE "80 86")/#&/g' ../cmake/Cuda.cmake && \ sed -ie 's/set(AMPERE "80 86")/#&/g' ../3rdparty/caffe/cmake/Cuda.cmake && \ make -j`nproc` && \ make install WORKDIR /openpose

One of the best cam only libraries (no depth sensor) I've seen is openpose, I ran it through a 360 camera and it was able to track body, face, and fingers really well even with spherical distortion from the 360 cam. example 360