@feufhd @yangxue0827 抱歉打扰了，我跑了val_crop将DOTA数据集分割，运行eval.py进行验证，但是出现了两个问题： 1.当我用demo_rh.py检测时，目视解译效果很不错，但是mAP效果不佳。旋转目标检测的结果如下: `cls : baseball-diamond|| Recall: 0.3277027027027027 || Precison: 0.6799065420560748|| AP: 0.2857233511210732 mAP is : 0.27839696674514086

rotation eval: Writing roundabout VOC results file Writing tennis-court VOC results file Writing swimming-pool VOC results file Writing storage-tank VOC results file Writing soccer-ball-field VOC results file Writing small-vehicle VOC results file Writing ship VOC results file Writing plane VOC results file Writing large-vehicle VOC results file Writing helicopter VOC results file Writing harbor VOC results file Writing ground-track-field VOC results file Writing bridge VOC results file Writing basketball-court VOC results file Writing baseball-diamond VOC results file cls : roundabout|| Recall: 0.05137844611528822 || Precison: 0.3942307692307692|| AP: 0.03896313329141163 cls : tennis-court|| Recall: 0.7942961165048543 || Precison: 0.9273822174991144|| AP: 0.7841216647894594 cls : swimming-pool|| Recall: 0.6034278959810875 || Precison: 0.4337298215802889|| AP: 0.4713228312417497 cls : storage-tank|| Recall: 0.08919098143236075 || Precison: 0.8359229334990678|| AP: 0.08359899690789072 cls : soccer-ball-field|| Recall: 0.4365482233502538 || Precison: 0.5|| AP: 0.37065461517016957 cls : small-vehicle|| Recall: 0.26025719534598896 || Precison: 0.44350029815146097|| AP: 0.14424862685392775 cls : ship|| Recall: 0.24484951836659447 || Precison: 0.7319509506751171|| AP: 0.20502145866938037 cls : plane|| Recall: 0.4970338303671637 || Precison: 0.8035251425609123|| AP: 0.47595213365825684 cls : large-vehicle|| Recall: 0.31162768312638367 || Precison: 0.6660152232051019|| AP: 0.23590015709743262 cls : helicopter|| Recall: 0.2857142857142857 || Precison: 0.42105263157894735|| AP: 0.2161832557262215 cls : harbor|| Recall: 0.1851032448377581 || Precison: 0.6090385198665453|| AP: 0.13665860486666898 cls : ground-track-field|| Recall: 0.4742268041237113 || Precison: 0.5302593659942363|| AP: 0.37650906662703715 cls : bridge|| Recall: 0.019583333333333335 || Precison: 0.27647058823529413|| AP: 0.006289862595004398 cls : basketball-court|| Recall: 0.37198795180722893 || Precison: 0.674863387978142|| AP: 0.33368834624831495 cls : baseball-diamond|| Recall: 0.3502252252252252 || Precison: 0.7300469483568075|| AP: 0.31682695380012843 mAP is : 0.2797293138362036 ` 2.当我用demo.py检测时，得到的_r.jpg和_h.jpg效果很差，基本上没有有效的检测框被标注出来。 另外提到一点，我在运行val_crop.py后对验证集val分割得到的图片和标注文件不能够完全一一对应，经过文件名的筛选，有139张图片没有对应的标注文件，也有139个标注文件没有对应的图片。因此我将val/images和val/annotation两个文件夹的文件做了一一对应的检查，然后运行eval.py得到了上面的结果。 我会继续排查mAP与目视解译效果不匹配的问题，也希望有人能给我一些帮助，谢谢：） 一点猜想：可能是我的标注文件有问题，所以用看起来效果很好的检测结果去eval的时候，好的检测结果匹配到了不正确的gt而降低？？

你好，repo的代码只是选取了resnet的最后一层作为特征，但是文章里说用了dense fpn的结构，请问这是为啥？没有开放嘛？ 文章是Position Detection and Direction Prediction for Arbitrary-Oriented Ships via Multiscale Rotation Region Convolutional Neural Network @yangxue0827

FYI: After you tag your images using labelme you run these scripts: train_crop.py val_crop.py

So the coordinates generated after chopping the images are wrong and in some cases totally wrong. Don't rely on this and write your own code for generating cropped images.

I compile the code and run inference in TF1.10.1, it reports

keep = tf.py_func(rotate_gpu_nms, inp=[det_tensor, iou_threshold, device_id], Tout=tf.int64) *** TypeError: cannot create weak reference to 'builtin_function_or_method' object

Could you help? Thanks

hello,

could you please shed some light on the input image resize that's performed in read_tfrecord.py-read_and_prepocess_single_img-image_preprocess.short_side_resize.

firstly can you please explain why this is done ?

additionally, can you also please explain why you have provided the default value of cfg.py: IMG_SHORT_SIDE_LEN = 800

Thanks, Omer

你好，想问一下lib/configs/cfgs中的这两个参数： IMG_SHORT_SIDE_LEN = 800 IMG_MAX_LENGTH = 1000 怎么用，因为我发现IMG_MAX_LENGTH参数貌似没有用到，所以想请教一些问题： 1）IMG_SHORT_SIDE_LEN可以随便设置吗？例如我有一批图片，最短边为200，必须设置为200吗？ 2）如果我要在512*512的图片上面进行训练，这两个参数应该如何设置？ 期待你的回复～

您好楼主！

我想问一下，我使用咱们R2CNN程序和训练好的权重文件108000.ckpt，在DOTA的val上跑结果。最终显示mAP只有0.38。然后各种调参数，重新从头训练到50w次了，mAP最多也只到0.48。同样的问题还有RRPN。 有点想不通这精度的差距能差在哪里，会跟机器和配置啥有关吗？还是我有什么地方没有注意到么？

望大哥解惑，谢谢了~~

I use 108000.ckpt, however I only got total mAP of 0.48 on DOTA val dataset. I didn't change anything of cfg parameters. I use the official DOTA_devkit to eval, it seems that some classes doesn't work well, like small vehicle only got 0.2 mAP.

C:\Users\Downloads\Compressed\R2CNN_Faster-RCNN_Tensorflow-master\libs\box_utils>python setup.py build_ext --inplace running build_ext Traceback (most recent call last): File "setup.py", line 132, in cmdclass={'build_ext': custom_build_ext}, File "C:\Users\Anaconda3\lib\distutils\core.py", line 148, in setup dist.run_commands() File "C:\Users\Anaconda3\lib\distutils\dist.py", line 966, in run_commands self.run_command(cmd) File "C:\Users\Anaconda3\lib\distutils\dist.py", line 985, in run_command cmd_obj.run() File "C:\Users\Anaconda3\lib\site-packages\Cython\Distutils\old_build_ext.py", line 186, in run _build_ext.build_ext.run(self) File "C:\Users\slcib\Anaconda3\lib\distutils\command\build_ext.py", line 340, in run self.build_extensions() File "setup.py", line 104, in build_extensions customize_compiler_for_nvcc(self.compiler) File "setup.py", line 77, in customize_compiler_for_nvcc default_compiler_so = self.compiler_so AttributeError: 'MSVCCompiler' object has no attribute 'compiler_so'

Hi, thanks for sharing your great work! I'm trying to use your model in text object detection. It predict very well for inclined object. However, With horizontal/vertical Object, the rotated box is bad as you can see images below Horizontal Object : https://drive.google.com/open?id=18yNNqqpmAEg6CAyqmXRGzvthAFpqOeFp Vertical Object : https://drive.google.com/open?id=1M2PGxyXaBeYwgIPLv71w9veh0rfbCNrd

Any suggestion on fixing this?

Best, Vu

I use ./lib/export_pbs to generate PB file, but It not work. the error as follow : ValueError: Shape must be rank 4 but is rank 5 for 'resnet_v1_101/Pad' (op: 'Pad') with input shapes: [1,1,?,?,3], [4,2]. can you tell me? how can I get the PB file, Thanks @yangxue0827

('class_list', 15) ('find image', 7) ('find label', 7) (0, 'read image', 'P2598.png') Traceback (most recent call last): File "train_crop.py", line 243, in clip_image(img.strip('.png'), img_data, box, 800, 800) File "train_crop.py", line 199, in clip_image save_to_xml(xml, subImage.shape[0], subImage.shape[1], box[idx, :], class_list) File "train_crop.py", line 116, in save_to_xml f = open(save_path,'w') IOError: [Errno 2] No such file or directory: 'rootdir/r2cnn/R2CNN_Faster-RCNN_Tensorflow/trail/save/labeltxt/P2598_0000_0256.xml'

File "txt2xml.py", line 124, in load_annoataion x1, y1, x2, y2, x3, y3, x4, y4 = list(map(float, line[:8])) ValueError: could not convert string to float: imagesource:GoogleEarth

The inference result is very different from what you expected. Which part should be modified?..... No action has been taken other than the guidelines presented in README.md.

Hi, recently I want to improve the result on my own dataset. Due to the reason that my result has many FP predicted bbox, so I try to adjust the parameter like show below:

1. FAST_RCNN_NMS_IOU_THRESHOLD = 0.5
   2.FAST_RCNN_IOU_POSITIVE_THRESHOLD = 0.8 3.FAST_RCNN_IOU_NEGATIVE_THRESHOLD = 0.5

When I was training , console popped out an error message:

2021-03-05 09:03:25.126176: F tensorflow/stream_executor/cuda/cuda_dnn.cc:430] could not convert BatchDescriptor {count: 0 feature_map_count: 512 spatial: 7 7 value_min: 0.000000 value_max: 0.000000 layout: BatchDepthYX} to cudnn tensor descriptor: CUDNN_STATUS_BAD_PARAM Aborted (core dumped)

Can someone tell me why,Thanks!

Modify: int(xy) --> int(float(xy))

Solving this issue: Traceback (most recent call last): File "data/io/DOTA/train_crop.py", line 244, in box = format_label(txt_data) File "data/io/DOTA/train_crop.py", line 136, in format_label [int(xy) for xy in i.split(' ')[:8]] + [class_list.index(i.split(' ')[8])] File "data/io/DOTA/train_crop.py", line 136, in [int(xy) for xy in i.split(' ')[:8]] + [class_list.index(i.split(' ')[8])] ValueError: invalid literal for int() with base 10: '1786.0'