University of Michigan - Ann Arbor

For the ImageNet Localization (LOC) task, per-class regression (PCR) learns a separate bounding-box regressor for each of the 1,000 object categories. The category-agnostic proposal network is replaced by a per-class RPN ending in two sibling $$1\times 1$$ convolutional layers: a 1,000-dimensional binary logistic classification layer that predicts class presence and a $$1000\times 4$$-dimensional regression layer that predicts coordinate offsets relative to translation-invariant anchor boxes. Because ImageNet images typically contain one dominant object, their highly overlapping region proposals produce nearly identical RoI-pooled features, reducing sample variance and hindering stochastic image-centric training. The framework therefore uses an RoI-centric R-CNN pipeline. During training, the top 200 proposals from the per-class RPN for ground-truth classes are cropped, warped to $$224\times 224$$ pixels, and sampled in mini-batches of 256 RoIs. During inference, the per-class RPN extracts the top 200 proposals for each predicted class, after which the R-CNN scores and regresses them. A single ResNet-101 model achieves 10.6% top-5 localization error on the ImageNet validation set, while an ensemble achieves 9.0% on the test set and first place in the ILSVRC 2015 localization challenge.

Per-Class Localization and RoI-Centric Training

Order the stages of the ResNet localization pipeline during inference from first to last.

Based on the behavior of region proposals in ImageNet, explain why image-centric training causes stochastic training to stall, and state the pipeline change adopted to overcome this issue.

Learn Before

Related