APPLIED COMPUTER VISION

Swan Lake Pose Estimation

I explored ViTPose for human pose estimation, adjusting model parameters to get the best visual output. Tested different confidence thresholds, keypoint visibility settings, and rendering styles to find the right balance between accuracy and visual clarity.

Kickup Counter

I built a system to detect and count football kickups from freestyle football videos.

Combined ViTPose for body pose estimation with YOLO for detecting the ball and player's feet. Calculated distances between foot keypoints and ball position, then analyzed ball velocity to determine when a kickup happened and increment the counter.

Used the Supervision library for annotations and labels, styling the output with custom colors and formatting.

Fashion Show Detection

I attempted object detection on runway models during a fashion show to detect and highlight specific clothing pieces from the collection.

Used Fashion CLIP and YOLO Fashion to find similar items based on reference images I provided.

Skateboarding Detection

Early exploration of YOLO models via Ultralytics. Applied object detection to skateboarding videos.

Traffic Annotation

I explored the Supervision library for annotations and labeling, applying detection and tracking to traffic footage while testing different annotation styles and label formats.

Snowboarding Detection

Applied YOLO via Ultralytics to snowboarding footage, detecting riders and equipment in varied outdoor conditions with snow, movement blur, and changing backgrounds.

Built with Python, YOLO, Supervision, PyTorch, and OpenCV. Models include RT-DETR, ViTPose, and Fashion CLIP. Video processing uses MoviePy, FFmpeg, and yt-dlp.