跳转至

红外目标检测实验结果收集

pytorch yolov3

Epoch Model P R mAP@0.5 F1 dataset
默认anchor v3tinyse 0.98 0.958 0.953 0.969 valid
默认anchor v3tinyse 0.933 0.895 0.878 0.913 test
默认anchor v3tinycbam 0.987 0.963 0.964 0.975 valid
默认anchor v3tinycbam 0.958 0.896 0.884 0.926 test
3a dt-16d-3a 0.982 0.958 0.95 0.97 valid
3a dt-16d-3a 0.922 0.894 0.876 0.908 test
6a dt-16d-6a 0.983 0.98 0.983 0.982 valid
6a dt-16d-6a 0.944 0.924 0.911 0.934 test
原始anchor v3tiny 0.971 0.978 0.973 0.974 valid
原始anchor v3tiny 0.931 0.899 0.884 0.915 test
6a dt-6a-cbam 0.983 0.978 0.974 0.98 valid
6a dt-6a-cbam 0.93 0.926 0.915 0.928 test
Epoch Model P R mAP F1 dataset
6a dt-6a-se 0.983 0.983 0.985 0.983 valid
6a dt-6a-se 0.936 0.93 0.919 0.933 test
6a dt-6a-dilated 0.957 0.971 0.964 0.964 valid
6a dt-6a-dilated 0.925 0.919 0.897 0.922 test
全部+se dt-6a-multi-se 0.98 0.976 0.973 0.978 valid
全部+se dt-6a-multi-se 0.913 0.941 0.926 0.927 test
backbone+se dt-6a-multi-se 0.983 0.978 0.973 0.98 valid
backbone+se dt-6a-multi-se 0.942 0.922 0.91 0.932 test
old:60 dense-v3-tiny 0.975 0.961 0.963 0.968 valid
old:60 dense-v3-tiny 0.945 0.908 0.89 0.926 test
修改好的 dense-v3-tiny 0.988 0.983 0.981 0.985 valid
修改好的 dense-v3-tiny 0.925 0.945 0.933 0.935 test
默认anchor yolov3 0.995 0.965 0.963 0.964 valid
默认anchor yolov3 0.953 0.88 0.876 0.915 test
Epoch Model P R mAP F1 dataset
6a dt-6a-spp 0.985 0.973 0.974 0.979 valid
6a dt-6a-spp 0.954 0.934 0.927 0.944 test
6a dt-6a-swish 0.983 0.983 0.981 0.983 valid
6a dt-6a-swish 0.936 0.93 0.923 0.933 test
pool->conv dt-6a-conv 0.978 0.976 0.974 0.977 valid
pool->conv dt-6a-conv 0.931 0.918 0.899 0.925 test
bbone-res10 res10-6a-spp 0.988 0.978 0.975 0.983 valid
bbone-res10 res10-6a-spp 0.934 0.923 0.913 0.928 test
Epoch Model P R mAP F1 dataset
9a csresnext50-panet-spp 0.98 0.98 0.984 0.98 valid
9a csresnext50-panet-spp 0.906 0.901 0.886 0.904 test
dt-8d-6a 0.99 0.946 0.943 0.968 valid
dt-8d-6a 0.942 0.885 0.872 0.912 test
dt-6a-bifpn 0.98 0.978 0.973 0.979 valid
dt-6a-bifpn 0.938 0.928 0.911 0.933 test
6a yolov3-tiny-prn 0.985 0.978 0.975 0.982 valid
6a yolov3-tiny-prn 0.936 0.922 0.914 0.929 test
6a yolov3-tiny-6a 0.973 0.98 0.984 0.977 valid
6a yolov3-tiny-6a 0.936 0.925 0.915 0.931 test
显存占用过大 dt-4d-6a 0.994 0.795 0.793 0.883 valid
显存占用过大 dt-4d-6a 0.978 0.728 0.715 0.835 test
Epoch Model P R mAP F1 dataset
dt-6a-prn-spp 0.993 0.983 0.985 0.988 valid
dt-6a-prn-spp 0.944 0.932 0.922 0.938 test
multi-scale+obj1 yolov3-tiny-6a 0.977 0.946 0.941 0.961 valid
multi-scale+obj1 yolov3-tiny-6a 0.936 0.88 0.865 0.907 test
过采样+FFM(os) dt-6a-spp-ffm 0.988 0.988 0.982 0.988 valid
过采样+FFM(os) dt-6a-spp-ffm 0.944 0.938 0.915 0.941 test
adam+obj(os) yolov3-tiny-6a 0.966 0.983 0.982 0.975 valid
adam+obj(os) yolov3-tiny-6a 0.907 0.915 0.899 0.911 test
过采样(os) dt-6a-spp 0.99 0.983 0.984 0.987 valid
过采样(os) dt-6a-spp 0.955 0.948 0.929 0.951 test
9a(os) yolov3-tiny-3l-9a 0.976 0.978 0.97 0.97 valid
9a(os) yolov3-tiny-3l-9a 0.932 0.93 0.917 0.931 test
过采样(os) dense-v3-tiny 0.993 0.985 0.983 0.989 valid
过采样(os) dense-v3-tiny 0.945 0.947 0.928 0.946 test
cbam branch(os) dt-6a-spp-cbam-branch 0.993 0.983 0.983 0.988 valid
cbam branch(os) dt-6a-spp-cbam-branch 0.949 0.934 0.914 0.942 test
过采样(os) enet-6a 0.993 0.993 0.993 0.993 valid
过采样(os) enet-6a 0.937 0.926 0.896 0.932 test
modifed-SE(os) dt-6a-multi-se 0.978 0.978 0.972 0.978 valid
modifed-SE(os) dt-6a-multi-se 0.939 0.938 0.921 0.939 test
aspp(os) dt-6a-aspp 0.99 0.99 0.992 0.99 valid
aspp(os) dt-6a-aspp 0.935 0.938 0.909 0.937 test
过采样(os) dense-v3-tiny-spp 0.993 0.993 0.993 0.993 valid
过采样(os) dense-v3-tiny-spp 0.952 0.949 0.932 0.951 test
pspnet中的结构(os) dt-6a-ppm 0.985 0.985 0.981 0.985 valid
pspnet中的结构(os) dt-6a-ppm 0.93 0.934 0.911 0.932 test
corner pool(os) dt-6a-corner-spp 0.988 0.985 0.982 0.987 valid
corner pool(os) dt-6a-corner-spp 0.937 0.938 0.917 0.937 test
dt-6a-dense-spp 0.985 0.983 0.983 0.984 valid
dt-6a-dense-spp 0.938 0.939 0.911 0.939 test
os dt-6a-spp2 0.99 0.985 0.982 0.988 valid
os dt-6a-spp2 0.949 0.945 0.925 0.947 test
dt-6a-spp-multi-se 0.99 0.978 0.972 0.984 valid
dt-6a-spp-multi-se 0.939 0.936 0.917 0.937 test
dt-6a-slim-spp 0.993 0.988 0.983 0.99 valid
dt-6a-slim-spp 0.937 0.928 0.899 0.933 test
LiteSeg aspp dt-6a-aspp 0.99 0.988 0.985 0.989 valid
LiteSeg aspp dt-6a-aspp 0.96 0.926 0.904 0.943 test
add1×1conv dt-6a-16d-up4 0.995 0.988 0.982 0.991 valid
1:2 dt-6a-16d-up4 0.94 0.93 0.914 0.935 test
baseline(os) yolov3-tiny原版 0.985 0.971 0.973 0.978 valid
baseline(os) yolov3-tiny原版 0.936 0.871 0.86 0.902 test
baseline yolov3-tiny原版 0.982 0.939 0.932 0.96 valid
baseline yolov3-tiny原版 0.96 0.873 0.869 0.914 test

SPP系列实验

Epoch Model P R mAP F1 dataset
baseline dt-6a-spp 0.99 0.983 0.984 0.987 valid
baseline dt-6a-spp 0.955 0.948 0.929 0.951 test
直连+5x5 dt-6a-spp-5 0.978 0.983 0.981 0.98 valid
直连+5x5 dt-6a-spp-5 0.933 0.93 0.914 0.932 test
直连+9x9 dt-6a-spp-9 0.99 0.983 0.982 0.987 valid
直连+9x9 dt-6a-spp-9 0.939 0.923 0.904 0.931 test
直连+13x13 dt-6a-spp-13 0.995 0.983 0.983 0.989 valid
直连+13x13 dt-6a-spp-13 0.959 0.941 0.93 0.95 test
直连+5x5+9x9 dt-6a-spp-5-9 0.988 0.988 0.981 0.988 valid
直连+5x5+9x9 dt-6a-spp-5-9 0.937 0.936 0.91 0.936 test
直连+5x5+13x13 dt-6a-spp-5-13 0.993 0.988 0.985 0.99 valid
直连+5x5+13x13 dt-6a-spp-5-13 0.936 0.939 0.91 0.938 test
直连+9x9+13x13 dt-6a-spp-9-13 0.981 0.985 0.983 0.983 valid
直连+9x9+13x13 dt-6a-spp-9-13 0.925 0.934 0.907 0.93 test
dt-6a-spp-acconv 0.988 0.983 0.983 0.985 valid
dt-6a-spp-acconv 0.947 0.945 0.92 0.945 test
1 3 4 5 dt-6a-aspp 0.993 0.985 0.98 0.989 valid
dt-6a-aspp 0.947 0.933 0.907 0.94 test
densenet201-bifpn valid
densenet201-bifpn test
yolov3-spp-9a 0.995 0.995 0.993 0.995 valid
yolov3-spp-9a 0.948 0.942 0.923 0.945 test
1,13|13,1|13,13 dt-6a-maxpoolone 0.99 0.988 0.982 0.989 valid
dt-6a-maxpoolone 0.956 0.941 0.928 0.948 test
并联 dt-6a-spp-13-se 0.942 0.943 0.922 0.943 test
dt-6a-spp-13-se 0.985 0.988 0.98 0.987 valid
dt-6a-spp-rfb 0.993 0.985 0.985 0.989 valid
dt-6a-spp-rfb 0.944 0.936 0.913 0.94 test

SPP 更改best标准后 系列实验

Epoch Model P R mAP F1 dataset
baseline dt-6a-spp 0.998 0.99 0.995 0.994 valid
baseline dt-6a-spp 0.948 0.938 0.919 0.943 test
直连+5x5 dt-6a-spp-5 0.993 0.983 0.982 0.988 valid
直连+5x5 dt-6a-spp-5 0.941 0.94 0.921 0.94 test
直连+9x9 dt-6a-spp-9 valid
直连+9x9 dt-6a-spp-9 test
直连+13x13 dt-6a-spp-13 valid
直连+13x13 dt-6a-spp-13 test
直连+5x5+9x9 dt-6a-spp-5-9 valid
直连+5x5+9x9 dt-6a-spp-5-9 test
直连+5x5+13x13 dt-6a-spp-5-13 valid
直连+5x5+13x13 dt-6a-spp-5-13 test
直连+9x9+13x13 dt-6a-spp-9-13 valid
直连+9x9+13x13 dt-6a-spp-9-13 test
dt-6a-add-spp 0.973 0.978 0.973 0.976 valid
dt-6a-add-spp 0.94 0.941 0.932 0.941 test
dt-6a-spp-11 0.99 0.988 0.982 0.989 valid
dt-6a-spp-11 0.95 0.944 0.93 0.947 test
yolov3-9a 0.99 0.995 0.994 0.993 valid
yolov3-9a 0.951 0.941 0.932 0.946 test
-max +up 4 dt-6a-16d-up4 0.99 0.985 0.983 0.988 valid
dt-6a-16d-up4 0.938 0.917 0.897 0.928 test
dt-6a-32d-up4 0.988 0.985 0.985 0.987 valid
dt-6a-32d-up4 0.943 0.922 0.904 0.932 test
filter=1:2 dt-6a-16d-up4 0.978 0.978 0.964 0.978 valid
dt-6a-16d-up4 0.924 0.935 0.913 0.93 test

Darknet

默认conf_thresh:0.2 iou_thresh=0.5 multi-scale

state model P R mAP F1 data
dt-6a-spp 0.97 0.97 0.963 0.97 valid
dt-6a-spp 0.92 0.89 0.8986 0.91 test
yolov3-tiny-6a 0.98 0.96 0.9696 0.97 valid
yolov3-tiny-6a 0.91 0.88 0.8651 0.90 test
dt-6a-prn-spp 0.97 0.97 0.9752 0.97 valid
dt-6a-prn-spp 0.91 0.90 0.8998 0.90 test
ignore=0.7 dt-6a-spp-fl 0.97 0.97 0.9755 0.97 valid
ignore=0.7 dt-6a-spp-fl 0.96 0.93 0.9294 0.94 test
ignore=0.3 yolov3-tiny-giou 0.97 0.97 0.9760 0.97 valid
ignore=0.3 yolov3-tiny-giou 0.90 0.89 0.8599 0.89 test
ignore=0.5 yolov3-tiny-giou 0.95 0.96 0.9636 0.95 valid
ignore=0.5 yolov3-tiny-giou 0.90 0.90 0.8681 0.90 test
ignore=0.3 dt-6a-spp-fl 0.95 0.99 0.9874 0.97 valid
ignore=0.3 dt-6a-spp-fl 0.89 0.92 0.9103 0.90 test

分背景测试

data num model P R mAP F1
trees 514 dt-6a-spp 0.947 0.998 0.992 0.972
sea_sky 39 dt-6a-spp 0.865 0.692 0.679 0.769
sea 25 dt-6a-spp 0.883 0.883 0.873 0.883
continuous_cloud_sky 512 dt-6a-spp 0.953 0.948 0.933 0.95
complex_cloud 515 dt-6a-spp 0.928 0.855 0.849 0.89
cloudless_sky 300 dt-6a-spp 0.997 0.987 0.984 0.992
architecture 483 dt-6a-spp 0.975 0.937 0.923 0.956

augData 训练

state model P R mAP F1 data
均衡数据 yolov3-tiny-6a 0.988 0.985 0.985 0.987 valid
均衡数据 yolov3-tiny-6a 0.95 0.935 0.919 0.942 test
sea sky pasted yolov3-tiny-6a 0.966 0.976 0.971 0.971 valid
sea sky pasted yolov3-tiny-6a 0.92 0.89 0.878 0.904 test
pasted 3times yolov3-tiny-6a 0.988 0.978 0.974 0.983 valid
pasted 3times yolov3-tiny-6a 0.946 0.916 0.905 0.931 test

均衡后的分背景测试

data num model P R mAP F1
trees 506 yolov3-tiny-6a 0.924 0.996 0.981 0.959
sea_sky 495 yolov3-tiny-6a 0.927 9.785 0.771 0.85
sea 510 yolov3-tiny-6a 0.923 0.935 0.893 0.929
continuous_cloud_sky 878 yolov3-tiny-6a 0.957 0.95 0.933 0.953
complex_cloud 561 yolov3-tiny-6a 0.943 0.833 0.831 0.885
cloudless_sky 1320 yolov3-tiny-6a 0.993 0.981 0.984 0.987
architecture 506 yolov3-tiny-6a 0.959 0.952 0.941 0.955

oversample+dt-6a-spp的分背景测试

data num model P R mAP F1
trees 506 dt-6a-spp 0.943 0.996 0.992 0.969
sea_sky 495 dt-6a-spp 0.927 0.785 0.777 0.85
sea 510 dt-6a-spp 0.897 0.909 0.842 0.903
continuous_cloud_sky 878 dt-6a-spp 0.951 0.942 0.912 0.946
complex_cloud 561 dt-6a-spp 0.94 0.886 0.882 0.912
cloudless_sky 1320 dt-6a-spp 0.993 0.981 0.985 0.987
architecture 506 dt-6a-spp 0.976 0.976 0.959 0.976
valid 391 dt-6a-spp 0.99 0.983 0.984 0.987
test 2388 dt-6a-spp 0.954 0.947 0.929 0.951

Anchor

3 anchor:

13, 18, 16, 22, 19, 25

6 anchor:

12,17, 14,17, 15,19, 15,21, 13,20, 19,24

9 anchor:

10,16, 12,17, 13,20, 13,22, 15,18, 15,20, 15,23, 18,23, 21,26

想法

如果Attention有效果,可以 Infrared Small Target Detection Network Based On Attetion。

1:减少下采样,先只用一个YOLO Layer 2:结合FPN,看能不能有精度提升(fuse方式创新)。 3:结合注意力机制看有没有性能提升(先尝试,如果有效可以结合)。

1:depwise conv / inverted bottleneck 2: xor/or/and netowrk 和 标准的network混合,可以去看yolo-nano,我们完全可以魔改一个自己的出来。

3:Focal loss未解决

data num auged
trees 506 506
sea_sky 45 495
sea 17 510
continuous_cloud_sky 878 878
complex_cloud 561 561
cloudless_sky 1320 1320
architecture 506 506

训练数据分析(综合测试结果来看)

背景类别 数量 特点 数据难度 测试mAP+F1 建议
trees 581 背景干净,目标明显,数量较多 0.99+0.97
cloudless_sky 1320 背景干净,目标明显,数量多 0.98+0.99
architecture 506 背景变化较大,目标形态变化较大,数量较多 一般 0.92+0.96 focal loss
continuous_cloud_sky 878 背景干净,目标形态变化不大,但个别目标容易会发生和背景中的云混淆 一般 0.93+0.95 focal loss
complex_cloud 561 目标形态基本无变化,但背景对目标的定位影响巨大 较难 0.85+0.89 focal loss
sea 17 背景干净,目标明显,数量极少 一般 0.87+0.88 生成高质量新样本,可以让其转为简单样本(Mixup)
sea_y 45 背景变化较大,且单张图像中目标个数差异变化大,有密集的难点,且数量少 困难 0.68+0.77 paste策
描述 模型 P R mAP F1 loss 数据集
rfb串联 dt-6a-spp-rfb 0.993 0.985 0.985 0.989 1.68 valid
dt-6a-spp-rfb 0.944 0.936 0.913 0.94 test
dt-6a-spp-15 0.988 0.983 0.983 0.985 1.87 valid
dt-6a-spp-15 0.954 0.933 0.918 0.943 test
dt-6a-spp-11-15 vaild
dt-6a-spp-11-15 test
dt-6a-spp-11-13-15 valid
dt-6a-spp-11-13-15 test