University of Michigan - Ann Arbor

For the ImageNet Localization (LOC) task, a per-class regression (PCR) approach is adopted, wherein bounding box regressors are learned individually for each of the 1,000 object categories. The proposal network is modified from category-agnostic proposals into a per-class RPN that ends with two sibling $1\times 1$ convolutional layers: a 1,000-dimensional binary logistic classification layer predicting class presence, and a $1000 \times 4$-dimensional regression layer providing coordinate offsets relative to translation-invariant anchor boxes.

While Fast R-CNN utilizes image-centric training, ImageNet images typically contain a single dominant object. Consequently, region proposals exhibit extensive spatial overlap and nearly identical RoI-pooled features, resulting in low sample variance that hinders stochastic training. To remedy this, the framework switches to an RoI-centric R-CNN pipeline. The top 200 proposals predicted by the per-class RPN for ground-truth classes are cropped, warped to $224\times 224$ pixels, and trained in mini-batches of 256 RoIs.

During inference, the per-class RPN extracts the top 200 proposals for each predicted class label, which the R-CNN then scores and regresses. A single ResNet-101 model achieves a 10.6% top-5 localization error on ImageNet validation, and an ensemble achieves 9.0% top-5 localization error on the test set, claiming first place in the ILSVRC 2015 localization challenge.

Per-Class Localization and RoI-Centric Training

Match each architectural component or parameter from the ImageNet per-class localization framework to its corresponding role.

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