Last week was the deadline for the ImageNet Large Scale Visual Recognition Challenge (ILSVRC 2015) large-scale object detection task. This is the primary challenge for image-based object detection. The challenge requires that you detect 200 classes of objects in a set of test images.
For each image, algorithms must produce a set of annotations (ci,si,bi)of class labels ci, confidence scores siand bounding boxes bi. This set is expected to contain each instance of each of the 200 object categories. Objects which were not annotated will be penalized, as will be duplicate detections (two annotations for the same object instance). The winner of the detection challenge is the team which achieves first place accuracy on the most object categories.
We won’t have the results of the challenge for a while now, but we achieved 52.7% (mean AP) on the validation set. This is as compared to Google’s 43.9% (44.5% on the validation set, last year’s winner) and CUHK’s 50.7% (the previous best performer).
This year’s results will inevitably be better than last year’s, and we may well not win as we are competing against teams with much better resources (and particularly GPUs) than we have access to. It’s a great result either way, however, and doubly impressive given that GPUs are very thin on the ground here.