Brand recognition in real-life photos
summary
A brand‑recognition pipeline uses pretrained VGG as a feature extractor and SVMs on Instagram crops to detect six brand pairs (Nike vs Adidas, Starbucks vs Costa, Heineken vs Carlsberg) with a VGG features + SVM classifier. Implementation uses Theano, Lasagne, Caffe. Detection scans 26 windows per image (sliding windows of different sizes) and runs from ~2 s down to ~20 ms per image after optimizations. Data augmentation and adding ~500 examples per class boost precision by a few percent each; layer choice and confidence cutoff also change results.
AI-written summary · about
chapters (25)
- 0:00Brand recognition in photos
- 1:05Instagram scale and brand challenges
- 3:14Why visual brand recognition matters
- 4:17Deep learning for image recognition
- 5:40Transfer learning with pre-trained networks
- 6:43Preparing images for neural networks
- 7:24Sliding window for logo detection
- 8:29Building dataset with logo classes
- 9:34Optimizing model speed
- 10:19How neural networks extract features
- 11:24Image preprocessing pipeline
- 12:05Training SVM with neural features
- 12:25Splitting train and test data
- 13:09Complete brand detection pipeline
- 14:14Improving results with more data
- 15:17Features from different network layers
- 16:02Adjusting probability thresholds
- 17:07Manually refining dataset
- 17:50Data augmentation in Python
- 18:32Normalizing data for performance
- 19:14Testing on real Instagram data
- 20:00Accuracy by brand class
- 21:56Final results and viability
- 22:31Q&A on image resizing
- 23:17Q&A on logo localization
related talks
41mOCR in the Wild World of Social Media
28mWhy I didn't use deep learning for my image recognition
46mImage Recognition with Deep Learning From prototype to production
33mMachine Learning in Visual Search
34mComputer Vision and NLP for Multi-Task Fashion Attrb. Modeling
59mAI in the Battle Against Fakes
46mML Use Cases at Think Big Analytics
01h 04mBuilding and Training your own Custom Image Recognition AI