Train Object Detection with Small Datasets
summary
Shows how to get usable detectors with a few hundred labeled images per class by combining three concrete tactics: remove near‑duplicates with hashing + latent clustering to choose diverse frames, extract 512‑D features from a pretrained backbone and sample uniformly per cluster, and use transfer learning (ImageNet backbone then COCO weights) so training converges faster and AP improves, especially at higher IoU. Implements preprocessing in PyTorch, COCO annotation export, training via MMDetection, and real experiments with YOLOv3 on insect subsets vs COCO baselines; the pipeline measurably raises AP and cuts annotation waste.
AI-written summary · about
chapters (19)
- 0:00Start
- 1:51Basics of Object Detection
- 4:50Refresher on Deep Learning
- 5:37Convolutions
- 8:16Gradient Descent
- 10:02Object Detection Models Are Built
- 10:20Two Stage Detectors
- 11:10The One Stage Detectors
- 16:04Iou Score
- 19:22Yolo V3
- 22:48Assembling a Diverse Dataset
- 27:56The Hashing Map
- 28:46How To Remove the Noise
- 31:06Gaussian Filter
- 41:57Cosine Distance
- 48:37Summary
- 54:38Average Precision
- 59:53Output of a Detection Model
- 1:04:08Cocoa Annotation Format
related talks
36mObject detection to Instance segmentation: Learn to apply several algorithms...
28mObject Detection using Transformers and CNNs
01h 29mHuman-like Visual Search Application with Small Data
01h 14mFoundations of Object Detection from an Expert's Overview
23mBuilding high-quality computer vision models using only a few examples
32mApplying Deep Learning for Object Detection from a Telecom Perspective
12mCascade models in computer vision to boost accuracy and performance
38mDeep Learning for detection on a phone