Building high-quality computer vision models using only a few examples
summary
Demo shows training a single-object detector from 21 pizza images with drawn bounding boxes and automatic hyperparameter selection. Production deployment exposes a managed hosting endpoint with AWS CLI and Python APIs. Research claims include learned high-dimensional embeddings instead of class means, better performance from regularized linear learners, a transductive fine-tuning loss that adds entropy minimization, semi‑supervised student‑teacher improvements, and fast "forgetting" via linear tailored expansions. Metrics reported are F1, precision and recall and per-image false positive/negative analysis.
AI-written summary · about
chapters (9)
- 0:00Introduction to AWS vision
- 2:44Recognition custom labels
- 3:40Workflow demo and training
- 7:52Inference and use cases
- 8:53Amazon Lookout for Vision
- 9:59Challenges of few-shot learning
- 12:39Few-shot learning techniques
- 17:09Advanced research contributions
- 20:51Summary and conclusion
related talks
01h 00mA gentle introduction to Deep Learning
18mCustomizable Computer Vision Expands Data Access Without Compromising Privacy
01h 15mTrain Object Detection with Small Datasets
01h 29mComputer Vision: An (Un?)Expected Journey, with Keras and Tensorflow
55mCore ML: Vision
40mMachines are Learning
01h 18mComputer Vision after the Victory of Data
35mComputer vision: British Clay