赞
踩
最近开始接触小目标检测的算法,直观想到的算法就是将可能存在小目标的图像区域进行合理的放大然后进行检测。偶然间发现与上述想法类似,原理简单却又有效的sahi算法而且能方便的与YOLO v5结合,正好拿来读一读。
论文题目:SLICING AIDED HYPER INFERENCE AND FINE-TUNINGFOR SMALL OBJECT DETECTION
论文地址
git地址
Fatih Cagatay Akyon, Sinan Onur Altinuc, & Alptekin Temizel (2022). Slicing Aided Hyper Inference and Fine-Tuning for Small Object Detection. In 2022 IEEE International Conference on Image Processing (ICIP). IEEE.
所谓小目标指的是目标成像尺寸较小,通常有两种定义方式:
在原始图片提取出一些patches,通过保持纵横比来调整大小(如图中红框所示),可以看作一种数据增强的操作。
如作者所说,必须注意的是,随着片尺寸减小,较大的对象可能不适合切片和相交区域,并且这可能导致对较大对象的较差检测性能。
基本过程可以描述为:将图片裁成一块块,然后再保持纵横比的同时调整每个patch的大小,分别进行预测,将预测结果合并回原始的大小同时与原始图像的全推断(FI)融合。然后用NMS(非极大值抑制)统一进行过滤。
对于小目标和大目标都可以取得不错的效果。
这里直接把作者的总结放在这里:
The proposed slicing aided hyper inference scheme can directly be integrated into any object detection inference pipeline and does not require pretraining. Experiments with FCOS, VFNet, and TOOD detectors on Visdrone and xView datasets show that it can result in up to 6.8% AP increase. Moreover, applying slicing aided fine-tuning results in an additional 14.5% AP increase for small objects and applying 25% overlap between slices results in a further 2.9% increase in AP. Training a network with higher resolution images through larger feature maps result in higher computation and memory requirements. The proposed approach increases the computational time linearly, while keeping memory require- ments fixed. Computation and memory budgets can also be traded-off by adjusting the patch sizes, considering the target platform. In the future, instance segmentation models will be benchmarked utilizing the proposed slicing approach and different post-processing techniques will be evaluated.
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。