Summary

Adds a second video tracking backend, sam3cpp_video, powered by PABannier/sam3.cpp — a C++14/ggml engine with no PyTorch dependency. This brings SAM 3 and
EdgeTAM to X-AnyLabeling for the first time and gives Apple Silicon / CPU-only users a video-tracking option without installing torch or sam-2.

The existing segment_anything_2_video backend is untouched.
This is to show some love to the mac people out there.

What's added

• New model class anylabeling/services/auto_labeling/segment_anything_3_video.py — graceful ImportError if the sam3cpp Python module isn't installed, matching the existing SAM 2 video class's
  pattern.

• 5 model entries in the AI model picker (auto-download from HuggingFace on first use):

  Model	Size	Notes
  sam3cpp_video-edgetam-q4_0	15 MB	mobile / lowest latency
  sam3cpp_video-edgetam-f16	27 MB	best EdgeTAM quality
  sam3cpp_video-sam3_visual-q4_0	289 MB	SAM 3 visual, quantized
  sam3cpp_video-sam3_visual-f16	946 MB	SAM 3 visual, full precision
  sam3cpp_video-sam3-q4_0	707 MB	SAM 3 full (text + visual), quantized

• Setup guide at examples/interactive_video_object_segmentation/sam3cpp/README.md — mirrors the existing SAM 2 example folder structure (uv venv, build sam3.cpp from source, install xlabel).

• Registry test tests/test_models/test_auto_labeling_registry.py — asserts no duplicate entries in any _*_MODELS list.

How a user installs it

The interaction model is identical to the existing SAM 2 video tutorial — see the new README for the full walkthrough. In short:

curl -LsSf https://astral.sh/uv/install.sh | sh                                                                                                                                                              
uv venv --python 3.10 .venv && source .venv/bin/activate                                                                                                                                                     
git clone --recursive https://github.com/PABannier/sam3.cpp                                                                                                                                                  
cd sam3.cpp && mkdir build && cd build && cmake .. && make -j                                                                                                                                                
# build sam3cpp Python bindings (see README — upstream PR pending)                                                                                                                                           
# clone & install X-AnyLabeling normally                                                                                                                                                                   

If sam3cpp is not importable, the model picker reports a clean error and the rest of the app keeps working.

Pending upstream work

The sam3cpp Python wrapper (pybind11) is being upstreamed to PABannier/sam3.cpp in a parallel PR. Once that lands, the bindings build collapses to a single CMake flag (-DSAM3_BUILD_PYBIND=ON) and the
README here will be updated. Link to that PR will be added to this description once it's open.

Verification

• All 5 HuggingFace .ggml URLs verified to resolve (HEAD 200).
• tests/test_models/test_auto_labeling_registry.py passes — no duplicate registry entries.
• bash scripts/format_code.sh produces no diff on the new file (black -l 79).
• Manual end-to-end on Apple Silicon: load sam3cpp_video-edgetam-q4_0 → click point on video frame → mask appears → "Run All" tracks across frames.
• The original segment_anything_2_video still loads (no regression).

Thanks for your contribution!
Please confirm that you agree to the Contributor License Agreement (CLA) by checking the box below:

• I have read and agree to the CLA.