Object Detection using Transformers and CNNs
summary
A car detector is fine-tuned on a drone dataset of ~6,000 training and ~500 validation images, using Detectron2 and Hugging Face tools and Python code snippets; experiments train Faster R-CNN with ResNet-50 + FPN and RetinaNet with focal loss and a transformer model DETR with ResNet-50 backbone. Faster R-CNN and RetinaNet reach similar COCO APs, RetinaNet wins on large cars, DETR shows good large-object AP but very poor small/medium AP and triggers OOMs unless images and batch sizes are reduced.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:36Dataset overview
- 2:21One-stage vs two-stage
- 3:44Faster R-CNN explained
- 6:55Feature pyramid networks
- 7:39Training Faster R-CNN
- 9:25RetinaNet with focal loss
- 11:32CNN detector performance
- 12:14Data augmentation
- 14:00Transformers in vision
- 16:46Self-attention in transformers
- 18:30Encoder-decoder variants
- 19:34Detection Transformer architecture
- 21:43Fine-tuning DETR
- 23:29DETR performance
- 24:36Improving DETR
- 25:40Future directions
related talks
01h 15mTrain Object Detection with Small Datasets
22mObject Detection with KerasCV
59mDeep Neural Networks for Video Applications
57mTensorflow, deep learning and modern convolutional neural nets, without a PhD
36mObject detection to Instance segmentation: Learn to apply several algorithms...
38mDeep Learning for detection on a phone
12mCascade models in computer vision to boost accuracy and performance
41mPractical Aspects of Applying Modern Computer Vision Technologies