sergeyk / timely

Timely Object Detection

This repository contains the working code for the NIPS 2012 submission on Timely Object Detection.

This is a private repository, and is superseded by the public code release (with supporting skvisutils and skpyutils repos) for the paper. However, this repo contains some code that was not published, and should be kept for archival purposes.

Files

• data/ is not tracked in the repository but should contain
  • results/: external repo for tracking generated results
  • VOC2007/: the 2007 VOC data
• synthetic/ is the code directory
• fastInf/ is external inference code

Overview

The project focuses on a multi-class detection policy. The overall motivation is the performance vs. time evaluation.

A. Single-class detector: window proposals

The evaluation method is plotting recall-#windows curves, keeping track of how long feature extraction and generating the windows take

• A.1. Jumping window proposals
  • mostly following Vijay's implementation
    • storing scale-invariant offsets instead of actual pixel values
  • ranked either by feature discriminativess or window parameter object likelihood (or combination)
• A.2. Sliding window proposals
  • parametrized in terms of min_width, stride, aspect_ratios, and scales.
  • these parameters are set from data in different ways
  • ranked either in fixed-order (scanning) or according to their object likelihood

B. Single-class detector: classification

• Chi-square (with Subhransu's trick) vs. RBF
• generate performance vs. time curves by timing how long it takes to process N window proposals from the first stage

C. Multi-class policy

Goal is to be able to learn a closed-loop policy

• Given a detector for each class, what is the most efficient way to search through the classes?
• What if there are multiple detectors per class (use the DPM detector as one of them, and ours as another)