The performance of UAVDT.

[jiangwenj02]

We re-trained the model with backbone resnet50 in UAVDT, and its performance as follows,

Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.156 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=500 ] = 0.302 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=500 ] = 0.139 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=500 ] = 0.005 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=500 ] = 0.113 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=500 ] = 0.251 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.308 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.379 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=500 ] = 0.379 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=500 ] = 0.042 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=500 ] = 0.339 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=500 ] = 0.455

It severe below the paper, how can i improve the performance? Can you release the guide of how to train the model in UAVDT?

Thank you very much.

[PuAnysh]

Hi, please show the mAPs of the coarse detector and mp-det on UAVDT dataset. I achieve 14% and 28.2% mAP respectively.

Since I have graduated and left school, I can't log in to the school server at present, so CKPT will be a little troublesome. If you have any problems in training, please communicate with me in the issue.

[jiangwenj02]

The results of coarse detector is bbox_mAP: 0.1370, bbox_mAP_50: 0.2440, bbox_mAP_75: 0.1410, bbox_mAP_s: 0.1070, bbox_mAP_m: 0.2130, bbox_mAP_l: 0.2520, bbox_mAP_copypaste: 0.137 0.244 0.141 0.107 0.213 0.252.

The results of mp-det is bbox_mAP: 0.2280, bbox_mAP_50: 0.3880, bbox_mAP_75: 0.2440, bbox_mAP_s: 0.0060, bbox_mAP_m: 0.2080, bbox_mAP_l: 0.3080, bbox_mAP_copypaste: 0.228 0.388 0.244 0.006 0.208 0.308.

The problem is seem in mp-det. I generate the dataset used in mp-det by following command, python UFPMP-Det-Tools/build_dataset/UFP_VisDrone2COCO.py \ ./configs/UFPMP-Det/UAVDT_coarse_det.py \ ./work_dirs/UAVDT_coarse_det/epoch_12.pth \ data/UAVDT/images/UAV-benchmark-M/ \ data/UAVDT/annotations/UAV-benchmark-M-Train.json \ data/UAVDT/images/UAV-benchmark-M-UFP/ \ data/UAVDT/annotations/UAV-benchmark-M-Train-UFP.json

Does it have any problem?

[PuAnysh]

This code is for the VisDrone dataset. Because UAVDT is annotated for tracking tasks, there are many ignore areas on the images (all instances that do not appear in the first frame will be identified with ignoring). I fill these areas with 0 when generating training datasets.

Here is another way to manually ignore the detection box in the ignored area by referring to the UAVDT evaluate code: https://drive.google.com/file/d/19498uJd7T9w4quwnQEy62nibt3uyT9pq/view.

Finally, Without filling ignored areas, the results of mp-det are also unsatisfactory. Please visualize the results of UFP algorithm to ensure the new training dataset is correct. I have provided the UAVDT dataset, which you can use directly!

[jiangwenj02]

Ok, i will try it. thank you very much.

PuAnysh @.***> 于2022年4月10日周日 23:14写道：

This code is for the VisDrone dataset. Because UAVDT is annotated for tracking tasks, there are many ignore areas on the images (all instances that do not appear in the first frame will be identified with ignoring). I fill these areas with 0 when generating training sets and test sets.

Here is another way to manually ignore the detection box in the ignored area by referring to the UAVDT evaluate code: https://drive.google.com/file/d/19498uJd7T9w4quwnQEy62nibt3uyT9pq/view.

Finally, Without filling ignored areas, the results of mp-det are also unsatisfactory. Please visualize the reuslts of UFP algorithm to ensure the new training dataset is correct.

— Reply to this email directly, view it on GitHub https://github.com/PuAnysh/UFPMP-Det/issues/2#issuecomment-1094294753, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFTV2O2ZGEZXYWPHCBQIDQLVELV33ANCNFSM5S3D4GKA . You are receiving this because you authored the thread.Message ID: @.***>

[mahilaMoghadami]

hello Did you reimplement and train this repository by yourself? I cant find mmcv and mmengine suitable version for this project. I got different errors while trying to train coarse-det.py. could you please help me? and share all packages version you used in your training?

[jiangwenj02]

hello Did you reimplement and train this repository by yourself? I cant find mmcv and mmengine suitable version for this project. I got different errors while trying to train coarse-det.py. could you please help me? and share all packages version you used in your training?

I can reimplement it. this repository is based on mmcv-full and doesn't need the mmengine. you can try to uninstall the mmcv and install the mmcv-full.

[mahilaMoghadami]

thank you so much could you please tell me your packages versions? mmcv-full and others?

[jiangwenj02]

i am sorry. i change my job and do not backup the environment.

获取Outlook for Androidhttps://aka.ms/AAb9ysg

---

From: mahilaMoghadami @.> Sent: Sunday, October 22, 2023 9:44:19 PM To: PuAnysh/UFPMP-Det @.> Cc: jiangwenj02 @.>; Author @.> Subject: Re: [PuAnysh/UFPMP-Det] The performance of UAVDT. (Issue #2)

thank you so much could you please tell me your packages versions? mmcv-full and others?

― Reply to this email directly, view it on GitHubhttps://github.com/PuAnysh/UFPMP-Det/issues/2#issuecomment-1774099663, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AFTV2O4ZI7GBWE7F24M3LALYAUPLHAVCNFSM5S3D4GKKU5DIOJSWCZC7NNSXTN2JONZXKZKDN5WW2ZLOOQ5TCNZXGQYDSOJWGYZQ. You are receiving this because you authored the thread.Message ID: @.***>

[peiqi999]

We re-trained the model with backbone resnet50 in UAVDT, and its performance as follows,

Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.156 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=500 ] = 0.302 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=500 ] = 0.139 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=500 ] = 0.005 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=500 ] = 0.113 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=500 ] = 0.251 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.308 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.379 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=500 ] = 0.379 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=500 ] = 0.042 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=500 ] = 0.339 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=500 ] = 0.455

It severe below the paper, how can i improve the performance? Can you release the guide of how to train the model in UAVDT?

Thank you very much.

@jiangwenj02 I have encountered AttributeError: module 'mmdet. models. backbones' has no attribute' MPHead '. How can I resolve it