Implemented a RunPod-based inference pipeline utilising DeepLabCut with SuperAnimal-TopViewMouse models. The system handles video upload, batch processing on NVIDIA A100 GPUs, and automated result download, enabling efficient processing of large video datasets without local hardware constraints.

Developed a comprehensive visualisation and analysis dashboard that generates per-video and per-individual behaviour summaries, chamber time pie charts, trajectory heatmaps, and stranger/novel object interaction metrics. The dashboard enables researchers to explore DeepLabCut outputs interactively.

Established a workflow to improve tracking accuracy through iterative refinement: extracting high-confidence frames from inference outputs, labelling using DeepLabCut's annotation tools, fine-tuning the SuperAnimal model on domain-specific data, and re-inferring videos with the improved model.

Implemented algorithms to quantify key behavioural metrics from pose estimation data: chamber occupancy time, nose proximity to stranger cup locations for social interaction scoring, movement patterns including distance travelled and velocity, and detection of exploratory behaviours such as sniffing, grooming, and freezing.

Integrated ezTrack as a complementary motion-based tracking system for rapid position/occupancy analysis and freezing behaviour quantification. This provides validation against DeepLabCut outputs and offers a fast alternative for specific analysis requirements.