[论文解读] Recent Advances in Object Detection in the Age of Deep Convolutional Neural Networks

Shivang Agarwal, Jean Ogier du Terrail+1 more|arXiv (Cornell University)|Sep 10, 2018

Advanced Neural Network Applications参考文献 457被引用 113

一句话总结

A comprehensive survey of deep CNN-based object detectors, detailing architectural designs (backbones, single- and double-stage detectors), training/inference practices, evaluation metrics, and future directions for extending detection to new modalities and constraints.

ABSTRACT

Object detection-the computer vision task dealing with detecting instances of objects of a certain class (e.g., 'car', 'plane', etc.) in images-attracted a lot of attention from the community during the last 5 years. This strong interest can be explained not only by the importance this task has for many applications but also by the phenomenal advances in this area since the arrival of deep convolutional neural networks (DCNN). This article reviews the recent literature on object detection with deep CNN, in a comprehensive way, and provides an in-depth view of these recent advances. The survey covers not only the typical architectures (SSD, YOLO, Faster-RCNN) but also discusses the challenges currently met by the community and goes on to show how the problem of object detection can be extended. This survey also reviews the public datasets and associated state-of-the-art algorithms.

研究动机与目标

• Summarize the evolution from hand-crafted to data-driven detectors and the impact of DCNNs on object detection performance.
• Analyze modern detector design choices (backbones, multi-scale representations, single- vs. double-stage frameworks) and training/inference strategies.
• Discuss evaluation metrics, datasets, and how detector performance is measured across major benchmarks.
• Identify current challenges (scale, rotation, domain adaptation, small objects, occlusion) and outline complementary ideas and new directions.
• Explore extensions of object detection to other modalities, constraints, and future goals such as interpretability and lifelong learning.

提出的方法

• Describe backbone network roles and the impact of classification backbones on detection performance.
• Explain single-stage and double-stage detector architectures and the role of region proposals and anchors.
• Discuss multi-scale detection, feature fusion, and top-down / bottom-up fusion strategies (e.g., FPN, RetinaNet).
• Summarize training components including losses, hyper-parameters, pre-training, and data augmentation.
• Outline inference strategies and post-processing, including IoU-based matching for evaluation metrics.
• Provide a synthesis of challenges and future directions for extending detectors beyond standard images.

实验结果

研究问题

• RQ1 What architectural choices and training strategies have driven performance gains in DCNN-based object detectors?
• RQ2 How do backbone design, multi-scale representations, and proposal mechanisms affect detection accuracy and speed?
• RQ3 What are the main challenges currently limiting object detection, and what complementary ideas show promise for overcoming them?
• RQ4 How can object detection be extended to other modalities (video, 3D, etc.) and under different constraints (weak supervision, few-shot, low power)?
• RQ5 What datasets and evaluation protocols shape the development and benchmarking of modern detectors?

主要发现

• Modern detectors largely build on fully-convolutional architectures with backbones adapted from image classification.
• RPNs and anchor-based frameworks underpin most state-of-the-art detectors, enabling end-to-end training and faster inference.
• Multi-scale feature representations and fusion (e.g., FPN) improve detection across object sizes and contexts.
• Performance on COCO, VOC, and related benchmarks is strongly influenced by backbone choice, data augmentation, and pre-training regimes.
• The survey highlights major challenges such as scale, domain shift, localization precision, and occlusion, and discusses complementary ideas like graph networks and contextual modeling.
• Extensions to detection tasks in video, 3D point clouds, and under constraints (weak supervision, few-shot, zero-shot, efficiency) are actively explored.