Updated Dependencies, Better Docker Support, and Segmentation Demo

[tim-win]

This PR introduces several significant improvements and updates to code surrounding the core YOLO-World project, addressing multiple issues and enhancing overall functionality and ease of use. When I finally got out of dependency hell, I decided to put down a ladder!

Key Changes

1. Dependency Updates:

   • Updated dependencies to match the latest recommended versions from issue #364, including torch >2.0.0 (phew!).
   • Upgraded to CUDA 12.1 to ensure compatibility with the latest GPU architectures and because thank god it works.

2. Docker Support:

   • Took the existing Docker demo system under my wing and cleaned it right up: it automatically handles the mm* dependency issues everyone has run into, as well as torch and others required for the demo.
   • Added a build_and_run.sh script for easy building and running of Docker containers with different model configurations, matching configs to models, so no one else needs the headache I have.

3. Segmentation Demo:

   • Added demo/segmentation_demo.py to showcase YOLO-World's open vocabulary segmentation capabilities. The guts of which was stolen shamelessly form @onuralpszr 's excellent hugginface space, https://huggingface.co/spaces/onuralpszr/YOLO-World-Seg, which did not work but showed me enough to get this running.
   • Integrated segmentation support into the Docker container, allowing for easy testing and demonstration of this feature.

4. Issue Resolutions:

   • This PR covers much of the work done in #419, bringing it up to date as of August 2024.
   • Implicitly fixes issues #279, #364, and #425.

5. Tested Configurations:

   • Verified functionality with pretrain-x-1280ft, which performs excellently.
   • Tested seg-l and seg-l-seghead configurations, which show good performance but really work well with my use case ( :/ )

Detailed Improvements

• Refactored the Dockerfile for better efficiency and clarity.
• Updated pyproject.toml and requirements files with pinned dependency versions.
• Minor changes to configuration files, there were some local paths that needed to be removed.
• Documentation the Docker-based demo workflow.

How to Use

Users can now easily run YOLO-World demos, including the new segmentation demo, using the provided Docker build system. For example:

./build_and_run.sh pretrain-x-1280ft  # For gradio object detection demo
./build_and_run.sh seg-l              # For segmentation demo

(note, while this is in MR, the fixes are not on master. So you have to replace this line in the dockerfile:

RUN git clone --recursive https://github.com/AILab-CVC/YOLO-World /yolo/

With this line:

RUN git clone --recursive https://github.com/tim-win/YOLO-World /yolo/

Hopefully this PR will save the people who come after me significant amounts of time. Feedback and further testing is welcome!

[tim-win]

Pinging @wondervictor as you may be able to review!

[surinder-ram]

unfortunately also this script is not free from bugs.

=> [final 6/6] RUN curl -o /yolo/weights/yolo_world_v2_x_obj365v1_goldg_cc3mlite_pretrain_1280ft-14996a36.pth -L https://huggingface.co/wondervictor/YOLO-World/resolve/main/yolo_world_v2_x 105.3s
=> exporting to image 22.4s
=> => exporting layers 22.4s
=> => writing image sha256:c55da7c781a19553466f44ef323893919c338711ea45264fbee8b1b514eacadc 0.0s
=> => naming to docker.io/library/yolo-demo:latest 0.0s
Running Docker container...
docker: invalid spec: :/weights/: empty section between colons.
See 'docker run --help'.