[McByte part 4] Add Cutie mask propagator

[tstanczyk95]

What does this PR do?

This PR adds a Cutie-based mask propagation component for McByte.

The new CutieMaskPropagator initializes Cutie memory from masks generated on a reference frame and propagates those masks to subsequent frames. The implementation integrates Cutie as an external dependency and follows the existing mask interfaces used by McByte.

Main additions:

• Added CutieMaskPropagator
• Added automatic Cutie weight download with checksum verification
• Added utilities for:
  • mask format conversion between McByte and Cutie
  • tracklet-to-object ID mapping
  • mask confidence computation
  • Cutie package/config discovery
• Added validation checks for mask shapes and tracklet-mask mappings
• Added unit tests covering helper utilities and propagator behavior
• Added a visual integration script for manual validation of SAM initialization and Cutie mask propagation

Type of Change

• New feature (non-breaking change that adds functionality)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:
Added unit tests covering:

• RGB image to Cutie tensor conversion
• Cutie probability tensor conversion to indexed masks
• Conversion between indexed and binary mask formats
• Non-overlapping mask generation
• Tracklet-to-object mapping construction
• Mask confidence computation
• Propagator state reset behavior
• Input validation during initialization

Manual validation:

• Tested CutieMaskPropagator initialization on CPU
• Tested CutieMaskPropagator initialization on CUDA
• Verified automatic checkpoint download
• Verified Hydra configuration loading
• Verified end-to-end SAM to Cutie propagation using the visualization script on a multi-frame sequence

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• [McByte as a whole is still in the development phase] I have updated the documentation accordingly (if applicable)

Additional Context

This PR is stacked on top of the SAM mask generator PR #441 and uses the masks produced by SAMBoxMaskGenerator for Cutie initialization.

It requires installation of torch, torch vision and SAM as specified in #441 (section Additional Context).

Furthermore, Cutie must be installed in the following manner (adapted steps from the original repo):

• Go to your parent folder of trackers.
• git clone https://github.com/hkchengrex/Cutie.git
• cd Cutie
• Open Cutie/pyproject.toml. Remove: 'cchardet >= 2.1.7', (whole line 43).
• pip install -e .

Tested on CPU and on GPU (CUDA 12.6) - separate Anaconda environments.

Visual test performed on SoccerNet-tracking:

• sequence SNMOT-116
• frames from from 000548.jpg to 000750.jpg (inclusive)
• oracle detections converted to xyxy format

Example test run for the visual test (you can also run it just with a single bounding box):

python visual_tests/visualize_cutie_mask_propagator.py \
  --image-path your_dataset_path/SoccerNet/tracking/test/SNMOT-116/img1/000548.jpg \
  --start-file 000548.jpg \
  --end-file 000750.jpg \
  --box 177,482,249,631 \
  --box 1241,569,1297,716 \
  --box 191,344,240,459 \
  --box 686,298,751,426 \
  --box 625,469,686,623 \
  --box 1382,378,1440,497 \
  --box 1220,480,1240,498 \
  --device cuda 

Analogously to SAMBoxMaskGenerator in #441, when CutieMaskPropagator is instantiated, it checks if the pre-trained Cutie weights file is already saved in the pointed location. If the model not saved, then it downloaded. Default location for the pre-trained model: trackers_root/models/cutie/cutie-base-mega.pth. Weighs from here.