Source-linked AI summary
Gaussian YOLOv3: An Accurate and Fast Object Detector Using Localization Uncertainty for Autonomous Driving
Jiwoong Choi, Dayoung Chun, Hyun Kim, Hyuk-Jae Lee
TL;DR
Autonomous-driving detectors must balance accurate localization with real-time speed, but existing approaches fall short on this trade-off. Gaussian YOLOv3 models YOLOv3 bounding boxes with Gaussian parameters and localization uncertainty, improving mAP by 3.09 on KITTI and 3.5 on BDD while maintaining real-time detection.
Problem
Autonomous-driving object detectors must achieve both high accuracy and real-time speed, but prior methods do not consistently satisfy both requirements.
Method
Gaussian YOLOv3 models YOLOv3 bounding-box coordinates with Gaussian parameters and uses predicted localization uncertainty during detection.
Results
The method improves mAP by 3.09 on KITTI and 3.5 on BDD compared with YOLOv3 while maintaining real-time detection above 42 fps.
Takeaways & Limitations
Gaussian YOLOv3 offers a stronger accuracy–speed trade-off for camera-based autonomous-driving object detection than the baseline and comparable-speed prior methods.
Takeaways & Limitations
Prior high-resolution detection methods cited by the paper did not exceed 30 fps, while a real-time alternative had low accuracy.
Abstract
from arXiv · showhide
The use of object detection algorithms is becoming increasingly important in autonomous vehicles, and object detection at high accuracy and a fast inference speed is essential for safe autonomous driving. A false positive (FP) from a false localization during autonomous driving can lead to fatal accidents and hinder safe and efficient driving. Therefore, a detection algorithm that can cope with mislocalizations is required in autonomous driving applications. This paper proposes a method for improving the detection accuracy while supporting a real-time operation by modeling the bounding box (bbox) of YOLOv3, which is the most representative of one-stage detectors, with a Gaussian parameter and redesigning the loss function. In addition, this paper proposes a method for predicting the localization uncertainty that indicates the reliability of bbox. By using the predicted localization uncertainty during the detection process, the proposed schemes can significantly reduce the FP and increase the true positive (TP), thereby improving the accuracy. Compared to a conventional YOLOv3, the proposed algorithm, Gaussian YOLOv3, improves the mean average precision (mAP) by 3.09 and 3.5 on the KITTI and Berkeley deep drive (BDD) datasets, respectively. Nevertheless, the proposed algorithm is capable of real-time detection at faster than 42 frames per second (fps) and shows a higher accuracy than previous approaches with a similar fps. Therefore, the proposed algorithm is the most suitable for autonomous driving applications.
1. Introduction
Autonomous-driving object detection must combine high accuracy with real-time speed, yet conventional detectors provide deterministic bounding boxes without localization uncertainty and can therefore produce mislocalization false positives. The paper motivates a YOLOv3-based approach intended to address these limitations while retaining fast multi-object detection.
- Motivation: Autonomous-driving detectors must accurately recognize road objects in real time to support safe control decisions and reduce response latency.Cameras are highlighted as cost-effective sensors that identify object types using texture and color features.
- Limitations: High-resolution inputs can improve detection accuracy but make previous object-detection methods unable to meet real-time requirements.The introduction specifically notes this limitation for methods using 512 × 512 or higher input resolution.
- Problem: Conventional detectors output deterministic bounding boxes without uncertainty, preventing them from identifying mislocalizations that cause false positives.The paper frames localization uncertainty as critical because known bounding-box coordinates alone do not indicate result reliability.
- Limitations: A prior uncertainty-prediction approach focused on a two-stage detector, lacked real-time operation, and retained bounding-box overlap problems, limiting its suitability for self-driving.These limitations motivate a method designed specifically for autonomous-driving applications.
- Contribution: The paper proposes a YOLOv3-based detector because YOLOv3 detects multiple objects in one inference and improves on YOLO and YOLOv2 through multi-stage detection.These speed and accuracy advantages are presented as making YOLOv3 suitable for autonomous-driving applications.
2. Background
YOLO detects multiple objects in one inference, making it faster than conventional methods, but grid-based processing causes localization errors. YOLOv3 improves accuracy while retaining speed, making it suitable for autonomous driving despite lower accuracy than two-stage detectors.
- YOLO: YOLO [19] outputs bounding-box coordinates, objectness, and class scores for multiple objects in one inference, providing faster detection than conventional methods.However, grid-unit processing produces large localization errors and low detection accuracy, limiting suitability for autonomous driving.
- YOLOv3: YOLOv3 [21] uses a deep convolutional network with residual skip connections and up-sampling with concatenation to improve accuracy and preserve fine-grained features for small-object detection.Its fully convolutional design uses only small 1 × 1 and 3 × 3 convolution filters, retaining fast detection speed.
- YOLOv3: YOLOv3 [21] offers a favorable accuracy–speed trade-off for autonomous driving and is widely used in autonomous-driving research, although it remains less accurate than two-stage detectors.Its speed is comparable to YOLO [19] and YOLOv2 [20], while its deeper architecture improves accuracy.
3. Gaussian YOLOv3
Gaussian YOLOv3 models each YOLOv3 bounding-box coordinate with Gaussian mean and variance, redesigns the bbox loss using negative log likelihood, and uses predicted localization uncertainty during detection. This preserves real-time efficiency while making detections more uncertainty-aware and robust to noisy training data.
- Gaussian bbox modeling: Gaussian YOLOv3 models each bbox coordinate as a Gaussian mean and variance, with the mean predicting the coordinate and the variance representing its uncertainty.The model outputs Gaussian parameters for tx, ty, tw, and th; center-coordinate means are constrained within each grid using sigmoid processing.
- Computational complexity: 0.04% is the computational-cost increase from single-Gaussian bbox modeling, requiring 99.04 × 10^9 FLOPs instead of YOLOv3’s 99 × 10^9 FLOPs.The Gaussian modeling applies only to bbox coordinates in the YOLOv3 detection layer, so the detection-speed penalty is extremely low.
- Loss redesign: The redesigned bbox loss uses negative log likelihood, allowing uncertainty to penalize inconsistent training data and focus learning on consistent data.Objectness and class losses remain unchanged, while the uncertainty-based loss attenuation improves robustness to noisy data.
- Uncertainty-aware detection: Gaussian YOLOv3 incorporates per-object localization uncertainty into detection confidence, lowering confidence when predicted bbox uncertainty is higher.Its detection criterion combines objectness, class confidence, and one minus the average uncertainty of the predicted bbox coordinates.
4. Experimental Results
Gaussian YOLOv3 is evaluated on autonomous-driving benchmarks using localization uncertainty, achieving higher accuracy than YOLOv3 and prior methods while retaining real-time speed. The experiments also examine uncertainty behavior and detection examples to assess improved TP and reduced FP.
- Uncertainty analysis: Localization uncertainty is evaluated against IoU for cars on KITTI and BDD validation sets, using averages within 0.1 IoU intervals.Figure 3 visualizes the relationship between IoU and the uncertainty predicted by the proposed algorithm.
- Benchmark comparisons: Gaussian YOLOv3 improves KITTI mAP by 3.09 over YOLOv3 while running at 43.13 fps, and exceeds RFBNet speed by 3.93 fps.Table 2 reports the KITTI validation comparison and defines the evaluation categories as easy, moderate, and hard.
- Benchmark comparisons: Gaussian YOLOv3 improves BDD mAP by 3.5 over YOLOv3 at 42.5 fps, exceeding RFBNet speed by 3.5 fps and accuracy by 3.9 mAP.The BDD comparison is reported on the test set in Table 3.
- Benchmark comparisons: On COCO, Gaussian YOLOv3 achieves AP 36.1 and AP75 39.0, respectively 3.1 and 4.6 higher than YOLOv3.These results extend the reported accuracy advantage beyond KITTI and BDD.
- Detection analysis: At the default detection threshold of 0.5, Gaussian YOLOv3 detects objects missed by YOLOv3 in KITTI and BDD examples, increasing true positives.The experiments use KITTI and BDD validation or test data, with KITTI containing car, cyclist, and pedestrian classes.
5. Conclusion
The paper proposes Gaussian YOLOv3 to address the accuracy–speed trade-off in camera-based autonomous-driving detection. Its Gaussian modeling, reconstructed loss function, and localization-uncertainty utilization yield higher accuracy at similar inference speed than previous studies.
- Conclusion: Gaussian YOLOv3 achieves the best trade-off between detection accuracy and operation speed for autonomous driving.The conclusion states that the proposed algorithm has higher accuracy than previous studies with similar fps.
- Conclusion: The method combines Gaussian modeling, loss-function reconstruction, and localization-uncertainty utilization to improve object detection.
- Conclusion: The proposed algorithm can significantly improve camera-based object detection systems and support the wider use of autonomous-driving applications.