SleapTAD: Temporal Action Detection for Sea Turtle Post-Hatchlings on Skeletal Features

Understanding the early behavioural development of sea turtle post-hatchlings is critical to resolving longstanding questions about the relative roles of active swimming and passive drift during the so-called “lost years”. Direct empirical observation of post-hatchling swimming is however constrained by the time-intensive nature of manual video analysis. We present SleapTAD, a temporal action detection system designed to automatically classify and localise two swimming behaviours (Powerstroke bouts and Dogpaddling) in untrimmed video recordings of loggerhead sea turtle post-hatchlings under controlled experimental conditions. SleapTAD combines a pose estimation backbone trained with Sleap on a purpose-built annotated skeleton dataset, a Transformer-based temporal encoder operating on normalised joint coordinates, and an ActionFormer detection head enabling deployment on consumer-grade hardware. We evaluate SleapTAD against two baselines: a TimeSFormer classifier with sliding-window temporal localisation, and AdaTAD, a state-of-the-art end-to-end temporal action detection model. On a dataset of 51 expert-annotated videos, SleapTAD achieves an average mAP of 40.89% across tIoU thresholds {0.1,0.4,0.5,0.6,0.7,0.8,0.9,0.95}, outperforming AdaTAD by approximately 11% points and TimeSFormer by over 20 percentage points, with the largest gains in the middle thresholds, while TimeSFormer retains an advantage at the strictest thresholds (tIoU≥0.9), suggesting complementary strengths between proposal-based and dense classification approaches. A systematic ablation study validates the key design choices of the architecture.