Detection Pipeline Refinements and Ensembling
Several architectural refinements built on top of residual representations provide substantial performance gains on detection benchmarks such as MS COCO and PASCAL VOC:
- Box Refinement: An iterative localization strategy where a new feature is pooled from the regressed bounding box to generate an updated classification score and bounding box. Combining these 300 regressed predictions with the initial 300 proposals, applying non-maximum suppression (NMS) with an IoU threshold of 0.3, and using box voting yields an improvement of approximately 2 points in mAP.
- Global Context: Global Spatial Pyramid Pooling (implemented as single-level RoI pooling across the entire image) is applied to the full-image convolutional feature map. This pooled global feature is passed through post-RoI layers and concatenated with the per-region feature before the final classification and regression layers, boosting mAP@.5 by about 1 point.
- Multi-Scale Testing: Feature maps are generated over an image pyramid with shorter side lengths s in {200, 400, 600, 800, 1000}. Features from two adjacent scales are pooled and merged via maxout layers, increasing mAP by more than 2 points.
By ensembling proposals and per-region classifiers across three networks, the Faster R-CNN system with ResNet-101 achieves 59.0% mAP@.5 and 37.4% mAP@[.5, .95] on COCO test-dev, winning first place in the COCO 2015 detection challenge.
0
1
Tags
Prep Sessions
Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
Ch.4 Residual Network Experiments and Applications - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
Object Detection and Localization using Residual Networks - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
Learn After
In the ResNet-based detection pipeline, how is the global context feature extracted and integrated with region-level features?
During multi-scale testing, features from two adjacent scales of the image pyramid are pooled and merged using maxout layers.
In the iterative box refinement process, what IoU threshold is applied during non-maximum suppression (NMS) when combining the 300 regressed predictions with the initial proposals?
Describe the ensembling strategy used with the ResNet-101 Faster R-CNN system, and state the benchmark performance results and placement achieved by this ensemble on MS COCO test-dev.