SpikeTrack

SpikeTrack runs all of Meta's SAM 3.1, from a text prompt to tracked masks, as one 12 MB Rust program on a Mac. We adapted it to find and track sperm cells in clinical microscope video. Why: sperm motility, a core part of male-fertility testing, is still judged largely by eye. Following dozens of small, fast, look-alike cells through video is hard, and fertility labs often cannot upload patient video to a cloud service. SAM 3.1's official code needs Python, PyTorch and a CUDA GPU. The existing Mac (Core ML) port runs only the tracker: it cannot take a text prompt, detect objects, or start new tracks. What we built: - All of SAM 3.1 in Rust: the CLIP text encoder, the DETR detector, the PE ViT backbone, the Object Multiplex tracker, and the detect, associate and track loop that starts and keeps tracks. It runs on our own Metal GPU engine (Ojas and tessl). Its heaviest parts (image trunk, necks, memory attention, memory encoder) run as half-precision Core ML programs exported from the checkpoint. It links only macOS system libraries and needs no Python at run time. - Checked against Meta's official code: every module matches (worst cosine 1.0000000, on CPU and on Metal). Image mode keeps the same detections on 20/20 frames. On 3 DAVIS clips, J&F equals the official code's. The full text-prompted tracking loop keeps the same tracked IDs on 32/32 frames. - Adapted to sperm: zero-shot, SAM 3.1 does not ground "sperm" (HOTA 0.000). We trained a LoRA adapter on the Mac (1.13M parameters, about 0.13% of the model) and applied a 768-scalar weight edit. Results on five held-out VISEM-Tracking videos: - Frame 0 detection: 95 of 112 heads, 83.7% F1. - Tracking over 1,500 frames: HOTA from 0.155 to 0.570. - Everyday negative prompts: 0 false positives out of 600. - Speed: about 1.0 s per frame on a dense video (M5 Pro), down from 2.9 s. It is not real time yet. Every number links to a saved run log in the repository. SpikeTrack is a research tool, not a medical device.