高效关键点检测：CornerNet Lite提升速度与精度

版权申诉

4 浏览量更新于2024-07-20 收藏 13.33MB PDF 举报

"CornerNet Lite：基于关键点的有效目标检测是一项前沿的AI技术，它在目标检测领域开辟了一种新的范式，即关键点检测，这摒弃了传统的锚定框方法，提供了一个更为简洁的检测框架。CornerNet，作为这种新方法的代表，已经在单级检测器中达到了最先进的性能，但其高精度伴随着较高的计算需求。在这个研究中，作者提出了CornerNet Lite，它是对原CornerNet的两大高效变体——CornerNet Saccade和CornerNet Squeeze的集成。CornerNet Saccade引入了一种注意力机制，通过减少对图像像素的全面处理，显著提升了处理效率，特别适用于那些对速度有较高要求的应用场景，例如在COCO数据集上，它的速度提升达到6.0倍，同时精度仅下降1.0%。另一变体CornerNet Squeeze则侧重于实时检测性能的提升，它采用了一种紧凑的网络架构（PactBackbone Architecture），旨在在保持准确性的同时提高YOLOv3等实时检测器的执行效率。在COCO数据集上，YOLOv3在30毫秒内的AP值从33.0%提高到了34.4%，这标志着基于关键点的检测方法在需要高效处理的应用中展现出了巨大潜力。 CornerNet Lite的出现不仅解决了高效目标检测的问题，还平衡了速度和精度，对于AI领域的实际应用具有重要意义，特别是在需要兼顾实时性和准确性的场景下。这项工作为后续研究者探索如何在关键点检测的基础上进一步优化性能和效率提供了新的方向。"

H. LAW, Y. TANG, O. RUSSAKOVSKY, J. DENG: CORNERNET-LITE 3

achieves an AP of 34.4% on COCO at 30ms, simultaneously more accurate and faster than

YOLOv3 (33.0% at 39ms).

A natural question is whether CornerNet-Squeeze can be combined with saccades to im-

prove its efﬁciency even further. Somewhat surprisingly, our experiments give a negative

answer: CornerNet-Squeeze-Saccade turns out slower and less accurate than CornerNet-

Squeeze. This is because for saccades to help, the network needs to be able to generate suf-

ﬁciently accurate attention maps, but the ultra-compact architecture of CornerNet-Squeeze

does not have this extra capacity. In addition, the original CornerNet is applied at multiple

scales, which provides ample room for saccades to cut down on the number of pixels to pro-

cess. In contrast, CornerNet-Squeeze is already applied at a single scale due to the ultra-tight

inference budget, which provides much less room for saccades to save.

Signiﬁcance and novelty: Collectively, these two variants of CornerNet-Lite make the

keypoint-based approach competitive, covering two popular use cases: CornerNet-Saccade

for ofﬂine processing, improving efﬁciency without sacriﬁcing accuracy, and CornerNet-

Squeeze for real-time processing, improving accuracy without sacriﬁcing efﬁciency.

Both variants of CornerNet-Lite are technically novel. CornerNet-Saccade is the ﬁrst to

integrate saccades with keypoint-based object detection. Its key difference from prior work

lies in how each crop (of pixels or feature maps) is processed. Prior work that employs

saccade-like mechanisms either detects a single object per crop (e.g. Faster R-CNN [48])

or produces multiple detections per crop with a two-stage network involving additional sub-

crops (e.g. AutoFocus [38]). In contrast, CornerNet-Saccade produces multiple detections

per crop with a single-stage network.

CornerNet-Squeeze is the ﬁrst to integrate SqueezeNet with the stacked hourglass archi-

tecture and to apply such a combination on object detection. Prior works that employ the

hourglass architecture have excelled at achieving competitive accuracy, but it was unclear

whether and how the hourglass architecture can be competitive in terms of efﬁciency. Our

design and results show that this is possible for the ﬁrst time, particularly in the context of

object detection.

Contributions Our contributions are three-fold: (1) We propose CornerNet-Saccade and

CornerNet-Squeeze, two novel approaches to improving the efﬁciency of keypoint-based

object detection; (2) On COCO, we improve the efﬁciency of state-of-the-art keypoint based

detection by 6 fold and the AP from 42.2% to 43.2%, (3) On COCO, we improve both the

accuracy and efﬁciency of state-of-the art real-time object detection (to 34.4% at 30ms from

33.0% at 39ms of YOLOv3).

2 Related Work

Saccades in Object Detection. Saccades in human vision refers to a sequence of rapid eye

movements to ﬁxate different image regions. In the context of object detection algorithms,

we use the term broadly to mean selectively cropping and processing image regions (sequen-

tially or in parallel, pixels or features) during inference.

There has been a long history of using saccades [12, 42, 63] in object detection to speed

up inference. For example, a special case of saccades is a cascade that repeatedly selects a

subset of regions for further processing, as exempliﬁed by the Viola-Jones face detector [56].

The idea of saccades has taken diverse forms in various approaches, but can be roughly

剩余14页未读，继续阅读

电动汽车控制与安全

粉丝: 266
资源: 4186

高效关键点检测：CornerNet Lite提升速度与精度

Peppa_Pig_Face_Engine：一种简单的人脸检测和对齐方法，既简单又稳定

CornerNet-Lite.pdf

hand_landmark_full.tflite与hand_landmark_lite.tflite区别

pose_landmark_lite.tflite

目标检测最新涨点结构

Lite Pose模型详解

lite-model_movenet_singlepose_lightning_tflite_int8_4.tflite详细的部署在安卓上的代码，测试一张图片

lite-model_movenet_singlepose_lightning_tflite_int8_4.tflite详细的部署在安卓上的代码，测试摄像头视频流

mediapipe 站、坐、躺识别 pythom

图片打标签有哪些工具？

最新资源