Visual Situation Recognition
summary
Project claims a system that recognizes concrete situations (example: dog-walking) by combining convolutional deep nets for low-level features with a Copycat-style active symbol architecture called scituate. It uses a supplied concept network/ontology, multi-agent codelets, a temperature parameter to control stochastic search, Bayesian/graphical belief updates, and heatmap weak segmentation. A prototype exists and gives probabilistic classifications with a temperature-based confidence; demos work intermittently and generalization beyond toy situations remains unfinished.
AI-written summary · about
chapters (28)
- 0:00Introduction
- 0:29Marvin Minsky
- 0:55Easy things are hard
- 1:20Deep Neural Networks
- 2:37Object Recognition
- 3:12Situation Recognition
- 4:23Semantic Gap
- 7:20Goal
- 13:36Conceptual Slippage
- 15:24Copycat
- 16:14Microworlds
- 16:48Letter strings
- 21:43Multiagent approach
- 29:02Concept Network
- 32:13Heat Maps
- 35:14Vision and Beyond
- 38:00generative models
- 38:48video
- 42:06compatibility
- 43:55dynamics
- 44:57other modalities
- 45:37languages
- 47:14processing power
- 48:46learning features
- 49:36parallel terrorists can
- 51:18hard problems
- 52:19computers cant do
- 53:30bottleneck
related talks
11mSaFace: Towards Scenario-aware Face Recognition via Edge Computing System
01h 08mMastering Visual AI with Vision-Language Models & Advanced Evaluation Techniques
32mWhy Current AI Cannot Truly See the World
21mRecognizing Human features using Deep Networks
41mThe Future of Computer Vision: How Two Rubyists Are Changing The World
35mUnpacking Scattering Vision Transformer: A New Dawn in Computer Vision
27mHiding Objects from Computer Vision by Exploiting Correlation Biases
46mBuilding computer vision applications Q&A