Carven von Bearnensquash Paper Gestalt Carven von Bearnensquash
Background Peer review imperfect review process Growth in the volume of submissions, tripled over the last 10 years Less than ideal pool of reviewers General layout of a paper
Abstract Intuition: Quality of paper general layout of the paper Computer vision techniques to predict if the paper should be accepted Result: reject 15% of good papers, cut down the number of “bad papers” by more than 50%
Related work Unique work Text based – biased to certain terms: “boosting”, “svm”, “crf”, ignores rich visual information No previous work known
Approach Formulated as a binary classification task Training data set of example-label pairs, {(x1; y1); (x2; y2); ...(xn; yn)}, Xi: feature values for paper i, Yi: binary label, “good” or “bad” Goal: learn a function f: X {0, 1}
Approach Adaboost Select feature classifier with lowest error rate, increase weight of mis- classified data
Approach Empirical error is bounded by More math: Include Maxwell’s equations in the paper Equations improve paper gestalt
Features gradient, texture, color and spatial information LUV histograms, Histograms of Oriented Gradients and gradient magnitude.
Experiments - Data Acquisition Accepted papers from CVPR 2008, ICCV 2009, and CVPR 2009 as positive examples #1196 Workshop papers from these same conferences as an approximation as negative examples #665 Papers converted to images, resized and padded with blank pages. 25% testing and 75% training
Experiments - Assuming that rejecting 15% of good papers is acceptable, we can cut bad papers in half
Experiments “we’re not sure what this figure reveals” bar plots are particularly aesthetically pleasing
Experiments – good examples
Experiments – bad examples
Experiments – the paper itself The system reported a posterior probability of 88.4%, which reassured us that this paper is fit for the CVPR conference.
Conclusions The quality of a computer vision paper can be estimated well by basic visual features A real-time system to predict weather a paper should be accepted or rejected