The Search for a New Visual Search, Beyond Language
summary
Shutterstock deploys a launched system called composition aware search that claims faster, more precise image retrieval across 150 million assets by combining convolutional vision models, a non‑recurrent CNN language encoder, multimodal embeddings and an inverted multi-index for ANN lookup. Demonstrations include spatial queries like wine on the left cheese on the right, 2048‑dim image vectors, production A/B tests (some yield no CTR lift), and engineering on AWS Docker Kubernetes with TF Caffe and PyTorch.
AI-written summary · about
chapters (7)
- 0:00Start
- 2:30Areas for Deep Learning opportunity
- 8:40Deep Learning Frameworks
- 9:51Data Science - Language support
- 17:39Composition Aware Search (CAS)
- 18:56How does it work?
- 26:02Levels of Supervision
related talks
33mMachine Learning in Visual Search
28mVisual Search
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
33mArchitecting Scalable Multi-Modal Video Search
59mContextual search with vector search: exploring your options with open source tools
22mHarnessing image and text data for product matching
01h 26mBuilding an Image Search Engine with Generative Computer Vision Models
25mSemantic Search