Pose Tracking Project Planning Fundamentals
DeepLabCut (DLC) uses transfer learning from pretrained neural networks (ResNet, EfficientNet) to track user-defined body parts in video. Project planning involves three key decisions:
Body Part Selection: Choose anatomically distinct landmarks visible in all frames. More points increase labeling time linearly but improve kinematic reconstruction.
Frame Sampling: DLC recommends labeling 200–1,000 frames sampled uniformly across videos. The k-means algorithm in DLC selects maximally diverse frames. Over-labeling beyond 1,000 frames rarely improves accuracy.
Training Configuration: DLC uses data augmentation (10× by default) including rotation, scaling, and contrast changes. Training converges in 200k–500k iterations, taking 2–5 hours on a modern GPU.
Common Pitfalls in Pose Tracking Setup
Several factors degrade pose tracking quality beyond frame count:
• Low frame rate: Below 30 FPS, fast movements create motion blur and missing poses. Grooming and rearing in rodents require 60+ FPS
• Inconsistent lighting: Shadows and reflections shift apparent body part positions. Use diffuse, uniform illumination
• Labeling inconsistency: The single largest source of DLC error. Establish labeling conventions before starting and cross-validate between labelers
• Too few diverse poses: If training frames only show walking, DLC will fail on rearing, grooming, or turning. Sample frames from all behavioral states
• Skeleton topology errors: Incorrect edge connections produce crossed wireframes. Verify skeleton topology before training
• Over-reliance on augmentation: Augmentation helps but cannot compensate for fundamentally unrepresentative training sets
