ringringyi / dota_yolov2 Goto Github PK

I tried YOLO V2 on DOTA but the mAP is much lower than that in the paper. I was wondering if anyone can achieve the result and how to set the configuration for training. Thanks a lot.

用ResultMerge.py 合并结果，我已经用SWIG生成了 _ployiou.so , 为什么运行ResultMerge.py会出现错误： Importerror:dynamic module does not define module export function(PyInit__polyiou)
谢谢

I have run the pretrained model on test images from DOTA dataset. I have found few text files in results folder. Could someone tell me how we can visualize the results on test images ?

hello,
in forward_convolutional_layer sub-function, it calls "gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);", for my understanding, 'a' point to weight, 'b' point to input data, and 'c' is output. in general convolution calculation, each weight value in a conv kernel is multiplied with corresponding input data and calculate partial sum, that's to say, each weight value will multiply different value and do partial sum. but in gemm_nn sub-function, the same weight value 'A[i*lda+k]' multiply different 'B' value, and do partial sum, i am puzzled by this, could you help me ?

void gemm_nn(int M, int N, int K, float ALPHA, 
        float *A, int lda, 
        float *B, int ldb,
        float *C, int ldc)
{
    int i,j,k;
    for(i = 0; i < M; ++i){
        for(k = 0; k < K; ++k){
            register float A_PART = ALPHA*A[i*lda+k];
            for(j = 0; j < N; ++j){
                C[i*ldc+j] += A_PART*B[k*ldb+j];
            }
        }
    }
}

请现在valid_dota.sh所在的目录下创建results文件夹

你好，请问标出来的框的位置（即框的四个顶点的坐标）能在哪里找到呢？

您好！
想请问一下如何根据数据集更改anchors？
例如现在yolo-dota.cfg中给出的anchors是基于1024x1024得到的，假设我想训练yolov3，其中的anchors如何设置呢？

Couldn't open file: /home/yh/dota/dota_data/YOLO/test/test.txt
请问这个报错是来自哪个文件呢？

I may have a question related to your anchor size. How do you modify the anchor size in your cfg files? I found not all images after splitting in training are 1024*1024 so that it may need to resize images when training yolo. will it have effect on anchor sizes?

I just want to ask that, is there anyone who tried DOTA v1.0 with YOLOv3 or YOLOv4?

I am currently trying to train my model with YOLOv4-tiny but the mAP values are much less than the declared ones in the website. Could there be an incompatibility or other problem with the higher version YOLO algorithms?

我看了一下DOTA中有些例子的标注，有的bbox是歪着的。在Yolo的label里面只有x,y,w,h，没有bbox倾斜角度的信息，应该是默认bbox是正的。 我看了一下data_transform下的代码感觉也没有处理这个角度信息的地方。所以想问一下这个东西需要注意么，对训练结果影响如何？
谢谢

(ykw) MacdeMacBook-Air:darknet mac$ ./darknet detector train /Users/mac/Desktop/DOTA_YOLOv2-master/cfg/dota.data /Users/mac/Desktop/DOTA_YOLOv2-master/cfg/yolo-dota.cfg /Users/mac/Desktop/DOTA_YOLOv2-master/darknet19_448.conv.23 | tee bod-dota.txt
layer filters size input output
0 conv 32 3 x 3 / 1 1024 x1024 x 3 -> 1024 x1024 x 32 1.812 BFLOPs
1 max 2 x 2 / 2 1024 x1024 x 32 -> 512 x 512 x 32
2 conv 64 3 x 3 / 1 512 x 512 x 32 -> 512 x 512 x 64 9.664 BFLOPs
3 max 2 x 2 / 2 512 x 512 x 64 -> 256 x 256 x 64
4 conv 128 3 x 3 / 1 256 x 256 x 64 -> 256 x 256 x 128 9.664 BFLOPs
5 conv 64 1 x 1 / 1 256 x 256 x 128 -> 256 x 256 x 64 1.074 BFLOPs
6 conv 128 3 x 3 / 1 256 x 256 x 64 -> 256 x 256 x 128 9.664 BFLOPs
7 max 2 x 2 / 2 256 x 256 x 128 -> 128 x 128 x 128
8 conv 256 3 x 3 / 1 128 x 128 x 128 -> 128 x 128 x 256 9.664 BFLOPs
9 conv 128 1 x 1 / 1 128 x 128 x 256 -> 128 x 128 x 128 1.074 BFLOPs
10 conv 256 3 x 3 / 1 128 x 128 x 128 -> 128 x 128 x 256 9.664 BFLOPs
11 max 2 x 2 / 2 128 x 128 x 256 -> 64 x 64 x 256
12 conv 512 3 x 3 / 1 64 x 64 x 256 -> 64 x 64 x 512 9.664 BFLOPs
13 conv 256 1 x 1 / 1 64 x 64 x 512 -> 64 x 64 x 256 1.074 BFLOPs
14 conv 512 3 x 3 / 1 64 x 64 x 256 -> 64 x 64 x 512 9.664 BFLOPs
15 conv 256 1 x 1 / 1 64 x 64 x 512 -> 64 x 64 x 256 1.074 BFLOPs
16 conv 512 3 x 3 / 1 64 x 64 x 256 -> 64 x 64 x 512 9.664 BFLOPs
17 max 2 x 2 / 2 64 x 64 x 512 -> 32 x 32 x 512
18 conv 1024 3 x 3 / 1 32 x 32 x 512 -> 32 x 32 x1024 9.664 BFLOPs
19 conv 512 1 x 1 / 1 32 x 32 x1024 -> 32 x 32 x 512 1.074 BFLOPs
20 conv 1024 3 x 3 / 1 32 x 32 x 512 -> 32 x 32 x1024 9.664 BFLOPs
21 conv 512 1 x 1 / 1 32 x 32 x1024 -> 32 x 32 x 512 1.074 BFLOPs
22 conv 1024 3 x 3 / 1 32 x 32 x 512 -> 32 x 32 x1024 9.664 BFLOPs
23 conv 1024 3 x 3 / 1 32 x 32 x1024 -> 32 x 32 x1024 19.327 BFLOPs
24 conv 1024 3 x 3 / 1 32 x 32 x1024 -> 32 x 32 x1024 19.327 BFLOPs
25 route 16
26 reorg / 2 64 x 64 x 512 -> 32 x 32 x2048
27 route 26 24
28 conv 1024 3 x 3 / 1 32 x 32 x3072 -> 32 x 32 x1024 57.982 BFLOPs
29 conv 100 1 x 1 / 1 32 x 32 x1024 -> 32 x 32 x 100 0.210 BFLOPs
30 detection
mask_scale: Using default '1.000000'
Loading weights from /Users/mac/Desktop/DOTA_YOLOv2-master/darknet19_448.conv.23...yolo-dota
Done!

@ringringyi 您好！很感谢您的代码，训练过程中出现了一些问题：
1.按照说明加载了yolo-dota.cfg，只有单卡GPU,NVIDIA GT TITAN BLACK(算力3.5），训练1024大小的DOTA train数据集，因此设置：
batch=1
subdivisions=1
网络cfg中学习率是
learning_rate=0.00005
max_batches = 2000000
policy=constant
训练时已经加载了darknet19_448.conv.23。
训练约12h，共训练了109000张图然后停止了训练，因为得到的训练结果如下：
Region Avg IOU: 0.618628, Class: 0.671129, Obj: 0.088273, No Obj: 0.003471, Avg Recall: 0.666667, count: 3
Region Avg IOU: 0.350919, Class: 0.315243, Obj: 0.000529, No Obj: 0.002533, Avg Recall: 0.500000, count: 2
Region Avg IOU: 0.507692, Class: 0.390687, Obj: 0.003199, No Obj: 0.001610, Avg Recall: 0.600000, count: 5
Region Avg IOU: 0.586214, Class: 0.206978, Obj: 0.030189, No Obj: 0.002246, Avg Recall: 1.000000, count: 3
Region Avg IOU: 0.744804, Class: 0.428495, Obj: 0.018022, No Obj: 0.003115, Avg Recall: 1.000000, count: 2
Region Avg IOU: 0.219129, Class: 0.057150, Obj: 0.000013, No Obj: 0.002348, Avg Recall: 0.000000, count: 2

109280: 169.920197, 45.703640 avg, 0.000050 rate, 0.414891 seconds, 109280 images
109281: 26.782028, 43.811478 avg, 0.000050 rate, 0.408204 seconds, 109281 images
109282: 20.342304, 41.464561 avg, 0.000050 rate, 0.419111 seconds, 109282 images
109283: 24.117226, 39.729828 avg, 0.000050 rate, 0.408410 seconds, 109283 images
109284: 26.526783, 38.409523 avg, 0.000050 rate, 0.413556 seconds, 109284 images
109285: 15.803896, 36.148960 avg, 0.000050 rate, 0.419731 seconds, 109285 images
109286: 50.919559, 37.626019 avg, 0.000050 rate, 0.411063 seconds, 109286 images
109287: 0.020407, 33.865456 avg, 0.000050 rate, 0.420056 seconds, 109287 images
109288: 29.913115, 33.470222 avg, 0.000050 rate, 0.413893 seconds, 109288 images
109289: 28.218138, 32.945015 avg, 0.000050 rate, 0.425859 seconds, 109289 images
109290: 0.749020, 29.725416 avg, 0.000050 rate, 0.407215 seconds, 109290 images
109291: 0.018226, 26.754698 avg, 0.000050 rate, 0.415694 seconds, 109291 images
109292: 12.341995, 25.313427 avg, 0.000050 rate, 0.411180 seconds, 109292 images

我将输出结果分开了，第一部分去掉了报nan的行（大约有30%报nan)，cls和Obj在训练了10w张图片后仍然结果不高；第二部分的loss值和avg loss值在30以上。
2.停止训练后用10w次训练得到的weights对一张416大小的图像做了测试，效果不佳。下图是thresh=0.1的输出。

thresh>0.1时没有输出。
想请教这种状况是正常的吗，怎么可以调整？

我得到了split图的result文件，其中内容格式是imgname score xmin ymin xmax ymax。
问题一：请问我用DOTA_devkit/ResultMerge.py中的哪个函数完成merge？是mergebyrec吗？
问题二：我如何将最终的检测结果画到对应的图上？restored.showAnns是针对task1(定向边界框)的吗？,如何才可以画水平边界框结果？
谢谢

在cfg/yolo-dota.cfg文件中，设置anchors时应该是笔误吧：
anchors = 1.36,0.89, 1.99,2.14, 1.13,1.56, 2.66,2.92, 3.71, 4,31
最后一个应该是4.31，而不是4,31。
这个问题会造成检测报错。