[Paper Review] Object Detection based on Region Decomposition and Assembly

Seung‐Hwan Bae|arXiv (Cornell University)|Jan 24, 2019

Advanced Neural Network Applications14 references4 citations

TL;DR

This paper proposes a Region Decomposition and Assembly Detector (R-DAD) that improves object detection by decomposing object regions into smaller parts and fusing global and part-level features through learnable region assembly blocks. The method achieves state-of-the-art mAP of 81.2% on PASCAL VOC07, 82.0% on PASCAL VOC12, and 43.1% on MSCOCO18 without relying on data augmentation, feature pyramids, or multi-scale testing.

ABSTRACT

Region-based object detection infers object regions for one or more categories in an image. Due to the recent advances in deep learning and region proposal methods, object detectors based on convolutional neural networks (CNNs) have been flourishing and provided the promising detection results. However, the detection accuracy is degraded often because of the low discriminability of object CNN features caused by occlusions and inaccurate region proposals. In this paper, we therefore propose a region decomposition and assembly detector (R-DAD) for more accurate object detection. In the proposed R-DAD, we first decompose an object region into multiple small regions. To capture an entire appearance and part details of the object jointly, we extract CNN features within the whole object region and decomposed regions. We then learn the semantic relations between the object and its parts by combining the multi-region features stage by stage with region assembly blocks, and use the combined and high-level semantic features for the object classification and localization. In addition, for more accurate region proposals, we propose a multi-scale proposal layer that can generate object proposals of various scales. We integrate the R-DAD into several feature extractors, and prove the distinct performance improvement on PASCAL07/12 and MSCOCO18 compared to the recent convolutional detectors.

Motivation & Objective

• Address the degradation in detection accuracy caused by occlusions and inaccurate region proposals in CNN-based detectors.
• Improve feature discriminability by jointly modeling global object appearance and local part details.
• Enhance region proposal accuracy using a multi-scale proposal layer that rescales proposals without expensive image pyramids.
• Develop a learnable feature fusion mechanism that adaptively combines global and part features based on occlusion patterns.
• Achieve state-of-the-art performance without relying on external improvements like data augmentation, multi-scale testing, or model ensembling.

Proposed method

• Propose a multi-scale region proposal layer that re-scales region proposals from the RPN using different scaling factors to cover diverse object sizes.
• Apply region sampling with ratio-based balancing to handle class imbalance between object and non-object proposals.
• Decompose each object region into multiple smaller sub-regions to capture fine-grained part-level features.
• Extract CNN features from both the full object region and its decomposed sub-regions for holistic and part-level representation.
• Introduce Region Assembly Blocks (RABs) that learn semantic relations between part and global features through stage-wise aggregation of maximum responses.
• Combine global and part features using RABs to generate high-level, discriminative features for classification and localization.

Experimental results

Research questions

• RQ1 Can multi-scale region proposals improve detection accuracy without using image or feature pyramids?
• RQ2 Can joint modeling of global and part features enhance feature discriminability in occluded or small object detection?
• RQ3 Can a learnable feature fusion mechanism (RAB) adaptively weight reliable features under occlusion conditions?
• RQ4 Does the proposed R-DAD achieve state-of-the-art performance without relying on data augmentation, multi-scale testing, or model ensembling?
• RQ5 How robust is the R-DAD framework across different backbone networks and detection architectures?

Key findings

• R-DAD achieves a mAP of 81.2% on PASCAL VOC07 and 82.0% on PASCAL VOC12, outperforming previous methods without additional bells and whistles.
• On MSCOCO18, R-DAD achieves a mAP of 43.1% with ResNet-101, ranking highly on the official test-dev leaderboard.
• With multi-scale testing, R-DAD-v2 achieves 44.9% mAP on MSCOCO18 using ResNet-101, demonstrating strong performance under aggressive settings.
• Ablation studies confirm that both the multi-scale proposal layer and the RAB-based feature fusion contribute significantly to performance gains.
• Visual comparisons show that R-DAD reduces false positives and improves localization accuracy, especially for occluded objects.
• The method is robust and generalizable, as it improves performance across multiple feature extractors (ResNet-50, ResNet-101) and detection frameworks.