[Paper Review] Bag of Freebies for Training Object Detection Neural Networks

Zhi Zhang, Tong He+4 more|arXiv (Cornell University)|Feb 11, 2019

Advanced Neural Network Applications21 references146 citations

TL;DR

The paper studies training-time freebies for object detection that do not affect inference, showing up to 5% absolute mAP improvement on VOC/COCO across YOLOv3 and Faster R-CNN by stacking tricks like visually coherent mixup, label smoothing, cosine learning rate, synchronized BN, and random shapes.

ABSTRACT

Training heuristics greatly improve various image classification model accuracies~\cite{he2018bag}. Object detection models, however, have more complex neural network structures and optimization targets. The training strategies and pipelines dramatically vary among different models. In this works, we explore training tweaks that apply to various models including Faster R-CNN and YOLOv3. These tweaks do not change the model architectures, therefore, the inference costs remain the same. Our empirical results demonstrate that, however, these freebies can improve up to 5% absolute precision compared to state-of-the-art baselines.

Motivation & Objective

• Identify general training tricks that improve object detection performance across architectures without increasing inference cost.
• Quantify the impact of each training tweak on YOLOv3 and Faster R-CNN on Pascal VOC and MS COCO.
• Develop a cohesive set of guidelines for applying these tricks to single- and multi-stage detectors.

Proposed method

• Propose a visually coherent image mixup tailored for object detection that preserves spatial alignment.
• Evaluate training-time tweaks including learning rate scheduling, label smoothing, synchronized BatchNorm, and data preprocessing.
• Incrementally stack the tweaks to assess additive gains on single-stage and multi-stage detectors.
• Benchmark results on Pascal VOC and COCO using YOLOv3 and Faster-RCNN with fixed input scales and standard NMS.

Experimental results

Research questions

• RQ1 Do training-time freebies generalize across different object detection pipelines (single-stage vs. multi-stage)?
• RQ2 What is the individual and cumulative impact of mixup, label smoothing, cosine LR scheduling, synchronized BN, and other tweaks on mAP?
• RQ3 Can these training tricks close the gap to state-of-the-art baselines without changing network architectures?
• RQ4 Are gains consistent across datasets (Pascal VOC and COCO) and input resolutions?

Key findings

• Up to 5% absolute mAP improvement over baselines without changing architectures.
• Mixup tailored for object detection yields measurable gains (e.g., 3.43% cumulatively when stacked with other tricks on VOC).
• Data augmentation alone contributed a large portion of gains for single-stage detectors (≈16% in COCO results).
• Synchronized BatchNorm, random training shapes, cosine LR schedule, class label smoothing, and mixup stack to produce consistent improvements across YOLOv3 and Faster-RCNN.
• On COCO, BoF improves YOLOv3 by up to 5.4% absolute mAP at 320–608 input resolutions and raises Faster-RCNN by about 1–2% overall.