bupt-ai-cz / LLVIP

LLVIP: A Visible-infrared Paired Dataset for Low-light Vision
610 stars 65 forks source link

#11

Closed Hiram1026 closed 1 year ago

SantJay commented 2 years ago

您好,可以提供一下您看的那片文章吗?我们目前发布的数据集在CVF openaccess版本的基础上做过更新,在更新后数据集上跑出来的指标确实会高一些。

Hiram1026 commented 2 years ago

好的 https://github.com/DocF/multispectral-object-detection 顺便请问一下你们CVF openaccess的文章能提供一下吗?

SantJay commented 2 years ago

您提供的这篇文章确实用了CVF openaccess版本LLVIP上的旧数据。其实应该以我们GitHub上的新数据为准,新数据也同步在了Arxiv版本LLVIP上。

Hiram1026 commented 2 years ago

目前从你们homepage取得的应该是最新的吧,顺便请问一下新老数据有没有明显的区分方式

SantJay commented 2 years ago

我们从未发布过旧数据集,从homepage上获取的就是最新的:)。

SantJay commented 2 years ago

我们使用的显卡是Tesla T4,测试模型使用的last.pt,imgz=1280,batchsize = 8,模型结构yolov5l。

XueZ-phd commented 1 year ago

请问下你们训练时的配置,什么显卡,最后是用的best.pt吗?

我在yolo上跑,自己用的rtx3090 imgz=1280 batchsize=8,我跑出了下面的结果,这个结果较高,我有点怀疑我的实验结果 image 我是用的best.pt基本上在50个epoch内就能取到了。

image

我们使用的显卡是Tesla T4,测试模型使用的last.pt,imgz=1280,batchsize = 8,模型结构yolov5l。

我也有同样的问题。我想知道作者在arxiv v3版本中使用的yolov5l是什么版本的?正如 @Hiram1026 所言,不同版本的yolov5l获得的结果是不同的,规律基本上是越新的版本结果越好。 我将作者arxiv上3个版本中的结果汇总如下,yolov5不同版本的时间也汇总如下。按照作者发表的时间,我猜测使用的是YOLOv5L 5.0 version。为了能够与作者的结果进行公平对比,我想知道作者的YOLOv5L版本号是什么。 image

image

SantJay commented 1 year ago

我们Arxiv v3版本中使用的yolov5l版本为version5.0,各位可以参考一下。

XueZ-phd commented 1 year ago

我们Arxiv v3版本中使用的yolov5l版本为version5.0,各位可以参考一下。

感谢回复!

heitorrapela commented 1 year ago

I saw that here you have test other versions of yolo. I have these results with the yolov5l (version 7.0) (commit: bfa1f23) on llvip infrared trained with the hparams from this repo.

infrared single modality image

rgb single modality image