First qualitative observations on deep learning vision model YOLO and DETR for automated driving in AustriaA PREPRINT
Figure 1: The first generation YOLO architecture [18].
(adding an improvement of about mAP of about 4%). YOLO was using arbitrary boundary boxes, with v2 bounding
boxes based on anchor box types were proposed with defined offsets to these anchor boxes to maintain the generic
capabilities (improvement of the recall). Typically, objects like a standing person or a car have a defined box ratio. In
contrast to YOLO, YOLOv2 now predicts for each bounding box the classes and not for each cell
3
. By k-means anchor
box dimension clusters, the data-driven selection (compared to hand selection) of anchor boxes was achieved based on
IoU
. Furthermore, the offsets of the anchor boxes were constrained with distant from the cell centroid. Furthermore,
capabilities for fine-grained features and multi-scale training added to YOLOv2 performance. A detailed discussion can
be found in [
19
]. In YOLOv3 [
20
], multi-label classification was used, since some classes are not mutual exclusive
(person, pedestrian, child, ...). In doing so, the soft max operation is avoided and the classification loss is now based on
binary cross-entropy. It makes 3 predictions per location at different resolution levels. One prediction is carried out
at the last feature map layer, one that upsamples features from two layers back by two. And a third, by going back
another two and upsample it again by two. YOLOv3 gained significant capabilities of detecting small objects [
20
].
Additional improvments on the usability and functionality were added with version 5 [
12
] (integrates the anchor-free
and objectness-free split head) and version 8 [
13
]. Version 8 can also be used for instance segmentation, skeleton
prediction of a human pose and classification. To conclude, YOLO’s real-time capabilities and easy to handle model
architecture are crucial for rapid object detection in autonomous driving scenarios.
2.2 RT-DETR
DETRs have achieved remarkable performance in object detection tasks. Initially, the high computational cost limits their
practical usage. Especially, the post-processing with non-maximum suppression is beneficial with the computational
cost, preventing original DETRs from being a new state-of-the-art (SOTA) for real-time object detection. The RT-DETR
was developed to solve the problem of high computational cost, above-mentioned [
15
]. In [
15
], it was shown how the
IoU
-threshold for admissible bounding boxes varies remaining prediction bounding boxes for YOLOv5 and YOLOv8.
Based on the number of remaining prediction bounding boxes, the non-maximum suppression takes a significant
execution time (depending on the
IoU
-threshold hyperparameters) and motivates the use of DETRs, with an overview
of the architecture in Fig. 2. Firstly, the big picture of RT-DETR
4
is discussed. As described in [
15
], RT-DETR consists
of a backbone, a hybrid encoder and a transformer decoder with auxiliary prediction heads. The last three stages of the
backbone
{S3, S4, S5}
are fed as input into the encoder. The efficient hybrid encoder processes multiscale features by
a process decoupling intra-scale feature interaction (AIFI) and cross-scale feature-fusion module (CCFM). The details
of the hybrid encoder (removing redundant operations of existing encoders) can be found in [
15
]. After the encoder, the
results are processed
IoU
-aware query selection. This is important to have the focus on the most relevant objects in the
scene by avoiding non-relevant parts and therefore enhancing the detection accuracy. The IoU-aware query selection
constraints the model to produce high classification scores for features with high
IoU
scores and low classification
scores for features with low
IoU
scores during training [
15
]. Finally, the decoder predicts outputs to generate boxes
and confidence scores. This design reduces computational costs and allows for real-time object detection on accelerated
backends, outperforming other real-time object detectors (see Fig. 3.
3
Some grafical explaination can be found here.
4
Implementation of RT-DETR is found on github.com/lyuwenyu/RT-DETR
4
剩余17页未读，继续阅读
人工智能_SYBH
- 粉丝: 4w+
- 资源: 220
我的内容管理
收起
- 我的资源
快来上传第一个资源
我的收益 登录查看自己的收益
我的积分
登录查看自己的积分
我的C币
登录后查看C币余额
我的收藏 
我的下载 
下载帮助
会员权益专享
最新资源
- VMP技术解析：Handle块优化与壳模板初始化
- C++ Primer 第四版更新：现代编程风格与标准库
- 计算机系统基础实验：缓冲区溢出攻击(Lab3)
- 中国结算网上业务平台：证券登记操作详解与常见问题
- FPGA驱动的五子棋博弈系统：加速与创新娱乐体验
- 多旋翼飞行器定点位置控制器设计实验
- 基于流量预测与潮汐效应的动态载频优化策略
- SQL练习：查询分析与高级操作
- 海底数据中心散热优化：从MATLAB到动态模拟
- 移动应用作业：MyDiaryBook - Google Material Design 日记APP
- Linux提权技术详解：从内核漏洞到Sudo配置错误
- 93分钟快速入门 LaTeX：从入门到实践
- 5G测试新挑战与罗德与施瓦茨解决方案
- EAS系统性能优化与故障诊断指南
- Java并发编程：JUC核心概念解析与应用
- 数据结构实验报告：基于不同存储结构的线性表和树实现
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
