Source-linked AI summary
PP-YOLOv2: A Practical Object Detector
Xin Huang, Xinxin Wang, Wenyu Lv, Xiaying Bai, Xiang Long, Kaipeng Deng, Qingqing Dang, Shumin Han, Qiwen Liu, Xiaoguang Hu, Dianhai Yu, Yanjun Ma, Osamu Yoshie
TL;DR
Practical object detection requires both effectiveness and efficiency, but combining accuracy refinements can produce inconsistent results. This paper incrementally evaluates PP-YOLO refinements and combines effective ones, producing PP-YOLOv2 with 49.5% mAP and 69 FPS while discussing unsuccessful trials.
Problem
Practical object detectors must improve accuracy despite limited computation resources and insufficient software support, while maintaining inference speed.
Method
The paper incrementally evaluates existing PP-YOLO refinements, combines effective changes, and discusses refinements that fail across dataset scales.
Results
49.5% mAP and 69 FPS give PP-YOLOv2 a better speed–accuracy balance than similarly sized detectors such as YOLOv4-CSP and YOLOv5l.
Takeaways & Limitations
PP-YOLOv2 provides a practical high-performance detector, and the reported refinement-combination experience is intended to help developers and researchers build object detectors.
Takeaways & Limitations
Some refinements improve COCO minitrain but hinder performance on COCO train2017, and cosine learning-rate decay is sensitive to hyperparameters.
Abstract
from arXiv · showhide
Being effective and efficient is essential to an object detector for practical use. To meet these two concerns, we comprehensively evaluate a collection of existing refinements to improve the performance of PP-YOLO while almost keep the infer time unchanged. This paper will analyze a collection of refinements and empirically evaluate their impact on the final model performance through incremental ablation study. Things we tried that didn't work will also be discussed. By combining multiple effective refinements, we boost PP-YOLO's performance from 45.9% mAP to 49.5% mAP on COCO2017 test-dev. Since a significant margin of performance has been made, we present PP-YOLOv2. In terms of speed, PP-YOLOv2 runs in 68.9FPS at 640x640 input size. Paddle inference engine with TensorRT, FP16-precision, and batch size = 1 further improves PP-YOLOv2's infer speed, which achieves 106.5 FPS. Such a performance surpasses existing object detectors with roughly the same amount of parameters (i.e., YOLOv4-CSP, YOLOv5l). Besides, PP-YOLOv2 with ResNet101 achieves 50.3% mAP on COCO2017 test-dev. Source code is at https://github.com/PaddlePaddle/PaddleDetection.
1. Introduction
The paper targets practical object detection by improving YOLO-based accuracy without sacrificing inference speed. It evaluates refinements incrementally and reports PP-YOLOv2 as a faster, competitive detector.
- Object detection supports applications including self-driving cars, face recognition, and person re-identification.
- 1.3% mAP is PP-YOLOv2’s advantage over YOLOv5l at similar FPS, while the ResNet101 variant is 15.9% faster than YOLOv5x at comparable performance.
- Practical detectors must address limited computation resources and insufficient software support while narrowing YOLOv3’s accuracy gap with two-stage detectors.
- The study incrementally evaluates refinements because methods that improve accuracy independently may not remain effective when combined.
- 49.5% mAP and 69 FPS give PP-YOLOv2 a better balance of effectiveness and efficiency than similarly sized detectors such as YOLOv4-CSP and YOLOv5l.
2. Revisit PP-YOLO
The paper establishes a PP-YOLO baseline through specified preprocessing, architecture, and training procedures. PP-YOLO is an enhanced YOLOv3 using ResNet50-vd and multiple efficiency-preserving refinements.
- PP-YOLO is an enhanced YOLOv3 that replaces its backbone with ResNet50-vd and adds 10 performance-improving tricks with little efficiency loss.
- Pre-Processing: Mixup uses Beta(1.5, 1.5) weights, followed by several augmentations applied with probability 0.5.
- Pre-Processing: Input sizes are sampled evenly from 320 through 608 pixels, while RGB channels are normalized using ImageNet-style means and standard deviations.
- Training Schedule: Training uses SGD for 500K iterations with minibatches of 96 images distributed across 8 GPUs.
3. Selection of Refinements
The selected refinements modify feature aggregation, activation placement, input scale, and IoU-aware supervision. Together they target stronger multiscale representations and improved training behavior while preserving the backbone or efficiency where specified.
- Path Aggregation Network: PAN aggregates top-down information through a detection neck that builds high-level semantic feature maps across object scales.
- Mish Activation Function: Mish is applied in the detection neck rather than the backbone to preserve the pretrained backbone parameters.
- Larger Input Size: Larger input sizes preserve small-object information more easily but require reducing batch size because they consume more memory.
- Larger Input Size: The larger-input strategy samples sizes evenly from 320 to 768 pixels and uses 12 images per GPU.
- IoU Aware Branch: The IoU-aware branch replaces a soft-weight loss format with soft labels, computing loss only for positive samples.
- IoU Aware Branch: The IoU-aware loss uses t for anchor-to-ground-truth IoU, p for the branch’s raw output, and σ(·) for sigmoid activation.
4. Experiments
Experiments evaluate PP-YOLOv2 refinements incrementally on COCO, measuring accuracy and inference speed. PAN, larger inputs, and modified IoU-aware loss improve performance, while comparisons show strong speed–accuracy results against similarly sized detectors.
- Experimental setup: COCO train2017 contains 118k images across 80 classes, while COCO minival provides 5k evaluation images using standard mAP.The reported ablation-study inference measurements use FP32 model time and exclude result decoding and NMS.
- Incremental ablation study: PAN with skip connections and Mish increases performance from 45.1% mAP to 47.1% mAP, despite being slightly slower.This refinement is adopted in the final model because of its substantial accuracy gain.
- Incremental ablation study: 0.6% mAP is gained by increasing training and evaluation input size to 640 for comparison with YOLOv4 and YOLOv5.The change establishes a fairer input-size comparison with those detectors.
- Incremental ablation study: 0.6% mAP is gained by using larger input sizes with 12 images per GPU, outperforming the larger-batch alternative in the ablation.The final training schedule samples sizes from 320 through 768.
- Incremental ablation study: 49.1% mAP is reached after modifying the IoU-aware loss, with no efficiency loss.The modified loss keeps IoU-aware and IoU losses at the same order of magnitude during training.
- Comparison with other detectors: TensorRT with FP16 precision and batch size 1 increases PP-YOLOv2 speed by 54.6% for ResNet50 and 73% for ResNet101.These speed-up ratios are reported for Paddle’s inference engine.
5. Things We Tried That Didn’t Work
The authors test refinements on COCO minitrain and COCO train2017, finding that some improvements do not transfer across dataset sizes. They discuss these inconsistencies to identify refinements with more universal behavior and explain failures.
- COCO minitrain contains 25K images and uses 90K training iterations, enabling faster ablation analysis than the roughly 80-hour COCO train2017 setup.The learning rate is divided by 10 at iteration 60K, while other settings match COCO train2017 training.
- COCO minitrain improvements can conflict with COCO train2017 results, motivating analysis of refinements across different dataset scales.The authors seek refinements that remain useful across datasets and investigate why some approaches fail.
- Cosine learning-rate decay improves COCO minitrain but produces no positive effect on COCO train2017 after testing several hyper-parameter settings.The method is sensitive to the initial learning rate, warm-up steps, and ending learning rate.
- Backbone parameter freezing yields a 1 mAP gain on COCO minitrain but decreases mAP by 0.8% on COCO train2017.The authors speculate that the discrepancy may reflect the different sizes of the two training sets.
6. Conclusions
The paper updates PP-YOLO into PP-YOLOv2 by combining and evaluating multiple refinements. It targets a better speed–accuracy balance and practical deployment support for object detection.
- PP-YOLOv2 combines multiple updates to PP-YOLO and forms a high-performance object detector.The paper explores a collection of tricks and demonstrates their effectiveness through incremental evaluation.
- The resulting detector is presented as balancing speed and accuracy better than detectors such as YOLOv4 and YOLOv5.The stated scope is practical object-detection performance and deployment.
- PaddlePaddle support narrows the gap between model development and production deployment.