TI.Resume

Computer Vision · 2025–2026

Identifikasi Walet

Individual identification from real-world CCTV footage

A computer-vision system that distinguishes individual swiftlets in noisy CCTV footage through detection, embeddings, and identity classification.

  • YOLO
  • ResNet
  • InsightFace
  • PyTorch
  • OpenCV
ID / 0299.33%test accuracy

Measured evidence

test accuracy
99.33%
test EER
0.00256
ROC AUC
0.99997
PR AUC
0.99991
best test F1 score
0.99562

Problem worth solving

Distinguish individual birds whose appearance is highly similar in noisy, real-world CCTV frames.

Engineering decisions

  1. Created an end-to-end workflow for frame extraction, bounding-box labeling, model training, inference, and monitoring.
  2. Compared CNN, YOLO, ResNet, and InsightFace-based embedding approaches.
  3. Built an inference web application for inspecting embedding separation and prediction performance.

What changed

  • Reached 99.33% test accuracy and a best test F1 score of 0.99562 on the reported test set.
  • Recorded a 0.00256 test EER, 0.99997 ROC AUC, and 0.99991 PR AUC.