some question about train with dataset NPS

[Yipzcc]

Dear author, thank you very much for your contributions. I have a question that I would like to ask you.

when i train NPS with your model or other model like yolov5s, yolov5l, yolov5l-xs-tph, I always get unsatisfactory results, with very low AP values, as shown in the figure below.

Here are some of the training commands.

1.yolov5l (move out the self.temporaltransformers in yolo.py)
python train.py --img 1280 --adam --batch-size 8 --epochs 80 --data ./data/NPS_1.yaml --weights ./weights/yolov5l.pt
--hy runs_ori/train/NPS/image_size_1280_temporal_YOLO5l_5_frames_NPS_end_to_end_skip_0/hyp_skip0_frames1.yaml
--cfg models/yolov5l.yaml --name 1280_1gpu_b8_oriv5l_pre_1 --project ./runs/train/NPS --single-cls --device 0

2.yolov5l-xs-tph （move out the self.temporaltransformers in yolo.py)
python train.py --img 1280 --adam --batch-size 4 --epochs 80 --data ./data/NPS_1.yaml --weights ./weights/yolov5l.pt
--hy runs_ori/train/NPS/image_size_1280_temporal_YOLO5l_5_frames_NPS_end_to_end_skip_0/hyp_skip0_frames1.yaml
--cfg models/yolov5l-xs-tph.yaml --name 1280_1gpu_b8_oriv5ltph_pre_1 --project ./runs/train/NPS --single-cls

i have checked out my dataset labels and found the the distribution of label data is almost identical to what you have uploaded, as shown in the figure below. I believe there is no problem with the data. Could you please give me some suggestions on how to solve this issue?

[defrag-bambino]

Hello my friend,
where can I download NPS and FLDrones datasets?

Thanks

[tusharsangam]

@defrag-bambino Have you tried running the pretrained model first ?
NPS dataset can be downloaded from https://engineering.purdue.edu/~bouman/UAV_Dataset/
for annotations please follow https://github.com/mwaseema/Drone-Detection these are different than original annotations, please use these.
FL-drones can be obtained from their respective authors, I got it from my co-author or I can send over email

[tusharsangam]

I will be uploading a detailed README soon. Or Happy to help over a meeting.

[defrag-bambino]

Thanks!
A detailed README with instructions on how to train etc. would be ideal. Because ultimately I want to train on a custom dataset.

[zhuang5252]

亲爱的作者，非常感谢您的贡献。我有一个问题想问你。

当我使用您的模型或其他模型（如 yolov5s、yolov5l、yolov5l-xs-tph）训练 NPS 时，我总是得到不令人满意的结果，AP 值非常低，如下图所示。

以下是一些训练命令。

1.yolov5l（移出 yolo.py 中的self.temporaltransformers） python train.py --img 1280 --adam --batch-size 8 --epochs 80 --data ./data/NPS_1.yaml --weights ./weights/yolov5l.pt --hy runs_ori/train/NPS/image_size_1280_temporal_YOLO5l_5_frames_NPS_end_to_end_skip_0/hyp_skip0_frames1.yaml --cfg models/yolov5l.yaml --name 1280_1gpu_b8_oriv5l_pre_1 --project ./runs/train/NPS --single-cls --device 0

2.yolov5l-xs-tph （移出 yolo.py 中的 self.temporaltransformers） 蟒蛇 train.py --img 1280 --亚当 --batch-size 4 --epochs 80 --data ./data/NPS_1.yaml --weights ./weights/yolov5l.pt --hy runs_ori/train/NPS/image_size_1280_temporal_YOLO5l_5_frames_NPS_end_to_end_skip_0/hyp_skip0_frames1.yaml --cfg models/yolov5l-xs-tph.yaml --name 1280_1gpu_b8_oriv5ltph_pre_1 --project ./runs/train/NPS --单cls

我检查了我的数据集标签，发现标签数据的分布与您上传的数据几乎相同，如下图所示。我相信数据没有问题。你能给我一些关于如何解决这个问题的建议吗？

Hello, I also encountered this issue when using YOLOv5n, YOLOv5s, YOLOv5l, and YOLOv5l xs tph. My accuracy, recall, and map did not increase but decreased. Have you resolved this issue?

[Yipzcc]

github太慢了，我csdn博客烤荔枝的大猫，这个代码nps数据集序号和dataset里面不一致，dataset对nps数据集会➕1 ---- 回复的原邮件 ---- 发件人 ***@***.***> 日期 2024年05月16日 10:26 收件人 ***@***.***> 抄送至 ***@***.***>***@***.***> 主题 Re: [tusharsangam/TransVisDrone] some question about train with dataset NPS (Issue #10) 亲爱的作者，非常感谢您的贡献。我有一个问题想问你。 当我使用您的模型或其他模型（如 yolov5s、yolov5l、yolov5l-xs-tph）训练 NPS 时，我总是得到不令人满意的结果，AP 值非常低，如下图所示。 以下是一些训练命令。 1.yolov5l（移出 yolo.py 中的self.temporaltransformers） python train.py --img 1280 --adam --batch-size 8 --epochs 80 --data ./data/NPS_1.yaml --weights ./weights/yolov5l.pt --hy runs_ori/train/NPS/image_size_1280_temporal_YOLO5l_5_frames_NPS_end_to_end_skip_0/hyp_skip0_frames1.yaml --cfg models/yolov5l.yaml --name 1280_1gpu_b8_oriv5l_pre_1 --project ./runs/train/NPS --single-cls --device 0 2.yolov5l-xs-tph （移出 yolo.py 中的 self.temporaltransformers） 蟒蛇 train.py --img 1280 --亚当 --batch-size 4 --epochs 80 --data ./data/NPS_1.yaml --weights ./weights/yolov5l.pt --hy runs_ori/train/NPS/image_size_1280_temporal_YOLO5l_5_frames_NPS_end_to_end_skip_0/hyp_skip0_frames1.yaml --cfg models/yolov5l-xs-tph.yaml --name 1280_1gpu_b8_oriv5ltph_pre_1 --project ./runs/train/NPS --单cls 我检查了我的数据集标签，发现标签数据的分布与您上传的数据几乎相同，如下图所示。我相信数据没有问题。你能给我一些关于如何解决这个问题的建议吗？ Hello, I also encountered this issue when using YOLOv5n, YOLOv5s, YOLOv5l, and YOLOv5l xs tph. My accuracy, recall, and map did not increase but decreased. Have you resolved this issue? _20240516102429.png (view on web) _20240516102445.png (view on web) — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[zhuang5252]

@defrag-bambino Have you tried running the pretrained model first ? NPS dataset can be downloaded from https://engineering.purdue.edu/~bouman/UAV_Dataset/ for annotations please follow https://github.com/mwaseema/Drone-Detection these are different than original annotations, please use these. FL-drones can be obtained from their respective authors, I got it from my co-author or I can send over email

Hello, I would like to use the following FL Dornes dataset. Could you please help me contact them?

[tusharsangam]

@defrag-bambino try to reach out to any of these three authors; You should be able to find their emails from the given paper citation.
A. Rozantsev, V. Lepetit, and P. Fua, “Detecting flying objects using
a single moving camera,” IEEE transactions on pattern analysis and
machine intelligence, vol. 39, no. 5, pp. 879–892, 2016.

[zhuang5252]

尝试联系这三位作者中的任何一位;您应该能够从给定的论文引用中找到他们的电子邮件。 A. Rozantsev、V. Lepetit 和 P. Fua，“使用单个移动相机检测飞行物体”，IEEE Transactions on Pattern Analysis and Machine Intelligence，第 39 卷，第 5 期，第 879–892 页，2016 年。

Okey, thank you very much.