Learning Representations for Visual Search
summary
Describes converting eye fixations into a probability distribution of fixations by Gaussian‑blurring fixation points, trains models with distributional losses (KL, Jensen‑Shannon, chi‑square) and finds chi‑square works well, experiments use a ResNet backbone for pixelwise saliency prediction, introduces generalized max pooling to aggregate thousands of local descriptors down to one vector for image retrieval, and shows that GMP compresses representations (1k–10k → 1) with better retrieval accuracy than average or max pooling.
AI-written summary · about
chapters (7)
- 0:00Introduction and welcome
- 1:47Guest introduction and background
- 8:51Visual attention research overview
- 13:50Research trajectory in visual attention
- 29:00Learning visual representations
- 38:32Current research focus and future directions
- 41:02Closing remarks and thanks
related talks
38mWhere are we looking? Predicting human gaze using deep networks.
28mVisual Search
43mThe Search for a New Visual Search, Beyond Language
17mLearning Cooperative Personalized Policies from Gaze Data
33mMachine Learning in Visual Search
01h 26mBuilding an Image Search Engine with Generative Computer Vision Models
51mTwo Concise Representations for High-Dimensional Visual Data
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models