[Paper Review] EAutoDet: Efficient Architecture Search for Object Detection

Xiaoxing Wang, Jiale Lin+3 more|arXiv (Cornell University)|Mar 21, 2022

Advanced Neural Network Applications4 citations

TL;DR

EAutoDet proposes an efficient differentiable neural architecture search framework for object detection that jointly optimizes backbone and FPN architectures in just 1.4 GPU-days using kernel reusing and dynamic channel refinement, achieving 40.1 mAP at 120 FPS and 49.2 mAP at 41.3 FPS on COCO, surpassing prior NAS methods without ImageNet pre-training.

ABSTRACT

Training CNN for detection is time-consuming due to the large dataset and complex network modules, making it hard to search architectures on detection datasets directly, which usually requires vast search costs (usually tens and even hundreds of GPU-days). In contrast, this paper introduces an efficient framework, named EAutoDet, that can discover practical backbone and FPN architectures for object detection in 1.4 GPU-days. Specifically, we construct a supernet for both backbone and FPN modules and adopt the differentiable method. To reduce the GPU memory requirement and computational cost, we propose a kernel reusing technique by sharing the weights of candidate operations on one edge and consolidating them into one convolution. A dynamic channel refinement strategy is also introduced to search channel numbers. Extensive experiments show significant efficacy and efficiency of our method. In particular, the discovered architectures surpass state-of-the-art object detection NAS methods and achieve 40.1 mAP with 120 FPS and 49.2 mAP with 41.3 FPS on COCO test-dev set. We also transfer the discovered architectures to rotation detection task, which achieve 77.05 mAP$_{ ext{50}}$ on DOTA-v1.0 test set with 21.1M parameters.

Motivation & Objective

• To address the high computational and memory cost of existing NAS methods for object detection, especially due to large-scale datasets and complex models.
• To enable end-to-end, joint architecture search for both backbone and FPN modules without pre-training on ImageNet.
• To reduce GPU memory and training time while maintaining high search accuracy and model performance.
• To explore the potential of combining NAS with YOLO-style one-stage detectors, which are fast but hard to improve via standard NAS due to their intricate handcrafted design.

Proposed method

• Constructs a single supernet that jointly models both backbone and FPN modules using a differentiable search strategy.
• Introduces kernel reusing by sharing weights across candidate operations on the same edge, consolidating them into a single convolution to reduce memory and computation.
• Employs dynamic channel refinement to adaptively search optimal channel numbers during training, improving convergence and efficiency.
• Uses macro search space to independently optimize C3-blocks, bottlenecks, and FPN fusion blocks, enabling fine-grained search over operations, channels, and connections.
• Trains the supernet from scratch on MS-COCO without ImageNet pre-training, significantly lowering search cost.
• Applies a two-stream optimization: alternately updating architecture parameters and network weights for 50 epochs using SGD.

Experimental results

Research questions

• RQ1 Can a joint search of backbone and FPN architectures be performed efficiently on detection datasets without pre-training on ImageNet?
• RQ2 How can kernel reusing and dynamic channel refinement reduce GPU memory and training time in supernet training for object detection?
• RQ3 Can NAS on YOLO-style one-stage detectors outperform handcrafted YOLO models in accuracy while maintaining high inference speed?
• RQ4 What is the performance ceiling of NAS when applied directly to COCO detection with minimal search cost?
• RQ5 How generalizable are the discovered architectures to other detection tasks, such as rotation detection?

Key findings

• EAutoDet discovers object detection models in only 1.4 GPU-days on a single V100 GPU, significantly reducing the search cost compared to prior methods that require tens to hundreds of GPU-days.
• The discovered AutoYOLO-s model achieves 40.1 mAP at 120 FPS on COCO test-dev, outperforming YOLOv5-s (36.9 mAP) with fewer parameters and higher speed.
• The AutoYOLO-x model reaches 49.2 mAP at 41.3 FPS, surpassing YOLOv5-x (49.1 mAP) while using slightly fewer FLOPs and parameters.
• The method enables direct training from scratch on COCO without ImageNet pre-training, eliminating the need for costly pre-training and improving search efficiency.
• When transferred to rotation detection on DOTA-v1.0, the AutoYOLO model achieves 77.05 mAP50 with only 21.1M parameters, demonstrating strong generalization.
• The search space is highly expressive, enabling fine-grained search over operations, channel multipliers, and connections, with over 5.4×10^8 candidate connection types per FPN block.