Thursday, January 28, 2010

Pedestrian detection discussion of evaluation and datasets

P. Dollar, C. Wojek, B. Schiele and P. Perona. Pedestrian Detection: A Benchmark. CVPR, 2009.

An analysis of pdestrain detection techniques. Provides a broader view of different techniques (beyond HoG) but also interesting consideration regarding evaluation metrics, datasets, conditioning of training vs. test data, which are intersting beyond this particular recognition task.

2 comments:

  1. The thing that strikes me the most after reading this paper is how terrible the results are for pedestrian detection (for all compared algorithms). The best overall performer (MultiFtr) achieves a miss rate of "only" 82% at 1 FPPI. Even at the easiest experiment (near scale and no occlusion), the best performer achieves a miss rate of 39% at 1 FPPI. Indeed, the authors say it is a very challenging dataset, and given the results I am inclined to believe it.

    ReplyDelete
  2. True, but now I'm waiting for the Per-sequence evaluation. I'm sure incorporating temporal information will give less depressing numbers.

    ReplyDelete