Human-like Visual Search Application with Small Data
summary
Builds a Siamese Mask R-CNN for one‑shot instance segmentation, trained from COCO with 80 categories, uses pre‑trained checkpoints (large config ~1 week on 4 GPUs, small config ~4–5 days), requires ~8 GB dataset to run. Prototypes a FastAPI single deployment that loads the model once and serves CPU inference; it works end‑to‑end but runs slowly and produces false positives in cluttered scenes (pigeon example), so pruning/quantization and retraining on novel classes are recommended.
AI-written summary · about
chapters (19)
- 0:00Workshop introduction
- 2:39Workshop overview
- 3:20Speaker intro
- 8:17Computer vision recap
- 10:24Siamese Mask R-CNN
- 14:41Setting up the environment and dependencies
- 23:55Mask R-CNN deep dive
- 27:54Q&A on Siamese networks
- 34:57Siamese Mask R-CNN for small data
- 37:12Visual search concept
- 43:38Deployment and FastAPI
- 53:47FastAPI endpoint
- 1:00:06Testing with Swagger UI
- 1:06:28Running predictions
- 1:10:00Limitations and improvements
- 1:14:15Workshop wrap-up
- 1:18:25Deployment best practices
- 1:24:06Open source contributions
- 1:28:21Q&A and closing
related talks
36mObject detection to Instance segmentation: Learn to apply several algorithms...
01h 15mTrain Object Detection with Small Datasets
33mMachine Learning in Visual Search
40m"Introduction to Face Processing with Computer Vision|"
30mStarting deep learning projects without enough labelled data
18mCustomizable Computer Vision Expands Data Access Without Compromising Privacy
01h 29mComputer Vision: An (Un?)Expected Journey, with Keras and Tensorflow
23mBuilding high-quality computer vision models using only a few examples