Bachelor theses in Deep Learning: Submitted to an Armenian University
summary
Bachelor theses present an object‑detection pipeline for top‑down vehicle imagery using 32,000 labeled vehicles from six locations at ~15 cm/pixel so cars are about 20×20 px, train a RetinaNet with Feature Pyramid Network backbone (freeze backbone, train heads), achieve about 95–96% weighted precision/recall, run inference at ~2 seconds per frame on two GPUs, and expose failure modes (vent shafts, shadows, dense parking bias) plus mitigation ideas like cutmix, synthetic shadows, keypoint labels and pruning/quantization for speed.
AI-written summary · about
chapters (8)
- 0:00GitHub Club and thesis supervision
- 1:55Convolutional neural networks basics
- 7:57Vehicle detection from satellite images
- 12:53RetinaNet and feature pyramid networks
- 15:42Training and performance results
- 18:37Optimizing model speed and real-time
- 20:49Future work and data augmentation
- 22:16Q&A on quantization and optimization
related talks
41mPractical Aspects of Applying Modern Computer Vision Technologies
59mDeep Neural Networks for Video Applications
31mDeep learning for self-driving vehicle technology
01h 30mDeep Learning Techniques for Monocular Depth Estimation
58mDeep Learning in Real Life
29mRecognizing Traffic Lights With Deep Learning
28mChallenges & future trends of Deep Learning
19mAI Trends in Intelligent Transport Systems