Why It Pays to Study Psychology - Lessons from Computer Vision

Explore a 31-minute conference talk by Alyosha Efros from the University of California, Berkeley, delving into the intersection of psychology and computer vision. Discover how studying psychology can benefit computer vision research, examining concepts like texture analysis, scene classification, and qualitative 3D scene reasoning. Learn about the work of Béla Julesz, the father of texture, and explore techniques such as texton discrimination and histogram matching. Investigate object recognition as texture recognition and the importance of learning to handle complexity. Examine qualitative 3D surface inference, occlusion reasoning, and line labeling techniques. Gain insights into the challenges of geometrically coherent image interpretation and the capacity of visual long-term memory. Understand how this research has inspired work in self-supervised representation learning, highlighting the valuable lessons computer vision can draw from psychological studies.

Intro
When are two textures similar?
Texture Analysis
Béla Julesz, father of texture
Texton Discrimination (Julesz)
Texture Synthesis input image
Scene Classification (Renninger & Malik)
Texton Histogram Matching
Discrimination of Basic Categories
Scene Recognition using Texture (2001)
Object Recognition is just Texture Recognition
Need learning to handle complexity!
Qualitative 3D Scene Reasoning
Support
Position, Probability, Size
3D Spatial Layout
Our Main Challenge
Infer Most Likely Scene
Qualitative 3D surfaces
The World Behind the image
Qualitative vs. Quantitative 3D
Occlusion Reasoning is Necessary
Line Labeling [Clowes 1971, Huffman 1971; Waltz 1972; Malik 1986]
Are Junctions local evidence?
Recover Major Occlusions
3D Depth Cues for Occlusion
Geometrically Coherent Image Interpretation
The Problem with Labeling
Simple 3D renderings
Capacity of Visual Long Term Memory (Aude Oliva)
how far can we push the fidelity of visual LTM representation ? Same object, different states
Inspired my work in Self-supervised Representation Learning