【Visdrone数据集】Visdrone+YOLOv7结果记录2

VisDrone+yolov7

训练

• 命令

python -m torch.distributed.launch --nproc_per_node 4 --master_port 9527 train.py --workers 4 --device 0,1,2,3 --sync-bn --batch-size 4 --data data/VisDrone.yaml --img 1536 2016  --cfg cfg/training/yolov7.yaml  --name yolov7_split_adam --hyp data/hyp.scratch.p5.yaml  --multi-scale --adam

分辨率：1536
结果输出所在文件夹：yolov7_split_adam5

验证集

matlab计算

• 命令：

• 结果

python计算

• 命令：

python test.py --data data/VisDrone.yaml --img 3200 --batch 8 --device 0,1 --weights /disk2/lxs/yolov7/runs/train/yolov7_split_adam9/weights/best.pt --name yolov7_val --task val            

• 结果

val: Scanning 'datasets/VisDrone/val_list.cache' images and labels... 548 found, 0 missing, 0 empty, 0 corrupted: 100%|███████████████████████████████████████| 548/548 [00:00<?, ?it/s]
               Class      Images      Labels           P           R      mAP@.5  mAP@.5:.95: 100%|█████████████████████████████████████████████████████| 69/69 [02:16<00:00,  1.97s/it]
                 all         548       38759       0.677       0.606       0.632       0.395                                                                                            
          pedestrian         548        8844       0.729       0.715       0.752       0.398                                                                                            
              people         548        5125       0.731       0.567       0.619       0.274                                                                                            
             bicycle         548        1287       0.599       0.439       0.477       0.243                                                                                            
                 car         548       14064       0.844       0.895       0.916       0.675                                                                                            
                 van         548        1975       0.639       0.621       0.641       0.478                                                                                            
               truck         548         750       0.684       0.543       0.578       0.406                                                                                            
            tricycle         548        1045       0.596       0.519       0.522       0.319                                                                                            
     awning-tricycle         548         532       0.456       0.342       0.301       0.194                                                                                            
                 bus         548         251       0.793       0.709       0.784       0.579                                                                                            
               motor         548        4886       0.698       0.711       0.728       0.379                                                                                            
Speed: 212.8/7.0/219.9 ms inference/NMS/total per 3200x3200 image at batch-size 8                                                                                                       
Results saved to runs/test/yolov7_val24  

测试集

matlab计算

• 命令：

python detect.py --device 0,1 --weights /disk2/lxs/yolov7/runs/train/yolov7_split_adam9/weights/best.pt --save-conf --save-txt --source /disk2/lxs/yolov5/datasets/VisDrone/VisDrone2019-DET-test-dev/images --img-size 2016 --name yolov7_split_adam9_img_2016

• 结果

Evaluation Completed. The peformance of the detector is presented as follows.
Average Precision  (AP) @[ IoU=0.50:0.95 | maxDets=500 ] = 0.33%.
Average Precision  (AP) @[ IoU=0.50      | maxDets=500 ] = 0.54%.
Average Precision  (AP) @[ IoU=0.75      | maxDets=500 ] = 0.37%.
Average Recall     (AR) @[ IoU=0.50:0.95 | maxDets=  1 ] = 0.12%.
Average Recall     (AR) @[ IoU=0.50:0.95 | maxDets= 10 ] = 0.73%.
Average Recall     (AR) @[ IoU=0.50:0.95 | maxDets=100 ] = 1.69%.
Average Recall     (AR) @[ IoU=0.50:0.95 | maxDets=500 ] = 1.76%.

python计算

• 命令：

python test.py --data data/VisDrone.yaml --img 2560 --batch 32   --device 0,1 --weights /home/sedlight/lxs/yolov7/runs/train/yolov7_split_adam9/weights/best.pt --name yolov7_test --task test

• 结果

test: Scanning 'datasets/VisDrone/test_list.cache' images and labels... 1610 found, 0 missing, 0 empty, 0 corrupted: 100%|██████████████████████████████████| 1610/1610 [00:00<?, ?it/s]
               Class      Images      Labels           P           R      mAP@.5  mAP@.5:.95: 100%|███████████████████████████████████████████████████| 202/202 [03:41<00:00,  1.10s/it]
                 all        1610       75102       0.598        0.54       0.529       0.315
          pedestrian        1610       21006       0.644       0.555       0.563       0.242
              people        1610        6376       0.632       0.347       0.375       0.138
             bicycle        1610        1302       0.493        0.29       0.298       0.142
                 car        1610       28074       0.786       0.874        0.88        0.58
                 van        1610        5771        0.59       0.589       0.589       0.418
               truck        1610        2659       0.636       0.617        0.62       0.425
            tricycle        1610         530       0.377       0.528       0.371       0.223
     awning-tricycle        1610         599       0.463       0.322       0.306       0.194
                 bus        1610        2940       0.774       0.666        0.73       0.535
               motor        1610        5845        0.59       0.609       0.559       0.256
Speed: 117.0/3.6/120.6 ms inference/NMS/total per 2560x2560 image at batch-size 8
Results saved to runs/test/yolov7_test9

yolov7加CBAM注意力机制

python -m torch.distributed.launch --nproc_per_node 2 --master_port 9527 train.py --workers 4 --device 0,1 --sync-bn --batch-size 2 --data data/VisDrone.yaml --img 1536 2016  --cfg cfg/training/yolov7_att.yaml  --name yolov7_split_adam_att --hyp data/hyp.scratch.p5.yaml  --multi-scale --adam

运行了几个epoch 效果并不明显 下次可以继续训练
输出文件：yolov7_split_adam_att6

yolov7加CBAM注意力机制2