Cascade models in computer vision to boost accuracy and performance
summary
Two-stage pipeline uses a fast high-recall detector on full frames downscaled to 640×640 to propose candidates, then a second-stage classifier trained on detector crops (separate classes for cell phone, umbrella, bag, other) validates each proposal so the classifier is idle ~99% of the time. Reported result: precision increases noticeably while recall drops only slightly. Recommended tooling includes YOLOv8 for detection, EfficientNet variants for classification, and TensorRT (2–5×) for faster inference.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:22preface
- 0:48classification
- 1:05object detection task
- 2:44difficulties
- 4:23cascade example
- 6:19why does it work?
- 7:38metrics
- 7:55dataset
- 9:24deployment
- 10:45examples
- 11:58thank you!
related talks
38mDeep Learning for detection on a phone
36mObject detection to Instance segmentation: Learn to apply several algorithms...
46mML Use Cases at Think Big Analytics
01h 15mTrain Object Detection with Small Datasets
40mWhen to go deep in Computer Vision... and how
41mPractical Aspects of Applying Modern Computer Vision Technologies
22mObject Detection with KerasCV
24mBeyond the Frame: Using Computer Vision to Unlock the Potential of Video Data