Why I didn't use deep learning for my image recognition
summary
A deep learning prototype trained with a VGG16 pre-trained base on ~300 images across 46 routes reaches ~95%+ test accuracy but fails practical checks: heavy model, TensorFlow instability, mobile inference latency, GPU retrain cost when routes reset, and no safe "unknown" class. Swapping to content based image retrieval using ORB + brute force matcher (OpenCV) gives ~99% retrieval accuracy, fast on-device or backend matching, immediate rejection of irrelevant photos, trivial addition of new routes, and an actually deployable Climbicus solution.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 2:30Climbing route matching problem
- 5:19User flow and hold trick
- 7:23Deep learning approach and data
- 9:09Training the neural network
- 12:17Model performance and results
- 13:41Deploying the model in production
- 15:47Handling unexpected images
- 17:52Limitations of classification models
- 18:56Route resets and retraining costs
- 20:19Content-based image retrieval
- 21:44ORB and brute force matching
- 23:12Performance comparison
- 24:36Solving new images and classes
- 25:40Easier deployment and costs
- 26:44Key learning and simplicity
- 27:25Conclusion
related talks
32mApplying Deep Learning for Object Detection from a Telecom Perspective
33mMachine Learning in Visual Search
46mImage Recognition with Deep Learning From prototype to production
58mDeep Learning in Real Life
41mPractical Aspects of Applying Modern Computer Vision Technologies
12mCascade models in computer vision to boost accuracy and performance
46mML Use Cases at Think Big Analytics
41mKEYNOTE: Image processing, AI, and autonomous systems