@mjq11302010044 I finally managed to run it :) thank you for your great support. I do not have cuDNN installed ... that might be an issue with the memory required by the demo? I am using a fairly new GPU, NVIDIA QUADRO-M4000 with 8GB! How is it possible I am running out of Memory? Also, I've resized your images, but still...:

libprotobuf WARNING google/protobuf/io/coded_stream.cc:505] Reading dangerously large protocol message. If the message turns out to be larger than 2147483647 bytes, parsing will be halted for security reasons.

Loaded network /home/vale/masterarbeit/H_AOSTD/data/faster_rcnn_models/vgg16_faster_rcnn.caffemodel F1109 05:23:29.260838 22269 syncedmem.cpp:64] Check failed: error == cudaSuccess (2 vs. 0) out of memory

I guess I could optimize utilization with cuDNN. But is it necessary? Quite bad experience with cuDNN...

The following example produces IOU greater than 1.

b1 = np.array([[46.83, 44.03, 3.9, 1.63, 0]], dtype=np.float32)
b2 = np.array([[46.83, 44.03, 1.63, 3.9, 1.45]], dtype=np.float32)
rbbox_overlaps(b1, b2) = 1.35

The expected iou should be near 1. Is there something I'm missing here? Note that I changed the code to use angle in radians.

你好。 请问在读取VOC数据集的时候，倾斜角θ是如何读入的？ 我查看了datasets/pascal_voc.py/_load_pascal_annotation，跟faster-rcnn一样，读取的只有bndboxxmin, xmax, ymin, ymax，并没有倾斜角θ的读取。 所以，请问是如何输入GT框的倾斜角θ的？

When I run the code ./tools/rotation_demo.py, I get the following error:

Loaded network /RRPN/data/faster_rcnn_models/vgg16_faster_rcnn.caffemodel Memory need is 426752000 Memory need is 426752000 Memory need is 106752000 Memory need is 106752000 Memory need is 213504000 Memory need is 213504000 Memory need is 53376000 Memory need is 53376000 Memory need is 106752000 F1010 11:27:03.680461 8528 syncedmem.cpp:57] Check failed: error == cudaSuccess (2 vs. 0) out of memory *** Check failure stack trace: *** Aborted (core dumped)

I have tried all possible available solution but they were not able to resolve this. I am compiling with Cudnn 7 and cuda 9.0. I have downgraded both of them and the problem was still not solved. I am using GT 710 2GB. Is there a way anyone can help me here. I am not even sure that if this is a bug or genuinely a hardware limitation. So before I go and buy a new GPU I would appreciate your help.

@mjq11302010044 @idefix92

Helo @mjq11302010044. I'm trying to create a model on my own just like your trained model that you shared.

But the training speed is very low. The iteration speed is aroud 5 seconds and the iteration count is 490000.

Do you have any trick to speed up the training speed? What can be the least iteration count to get the demo works?

Is it possible to list the major changes to this version of caffe you use? I want know the potential issues/conflict while merging it with the newer version of caffe.

Hi, I tested your pretrained VGG 16 model on ICDAR2015 with no parameters change. The following is the result of icdar2015 official evaluation script. All these values are lower than what is claimed in the paper.

Calculated!{"recall": 0.6721232546942706, "precision": 0.7977142857142857, "hmean": 0.7295531748105567, "AP": 0}

the code of function convert_region in rotate_polygon_nms_kernel.cu and rbbox_overlaps_kernel.cu seems to expect anti-clockwise degree as param. Do I have an exact understanding?

I compared the result with OpenCV3's rotatedRectangleIntersection, and the result equals.

my computer motherboard seems to be not compatible with Linux OS(asus Z370), anyone has tips for compiling the lib so it can be runned in win10 ??? thank you!!!

rotate_roi_align_layer.cu(147): error: calling a constexpr host function("fmax") from a global function("RotateROIAlignForward") is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this.

你好，谢谢你们的很棒的工作，请问一下， 1、如果想复现论文里的ICDAR2015的精度，是直接按照readme运行吗，我按照readme里直接运行，和论文里差20个点？ 2、我看caffe代码里没有实现数据增强，论文里说加了数据增强，是你们没有加进去吗？ 3、我看论文里的训练策略是前20w次迭代lr 1e-3,后10w次1e-4，这个和代码里的训练策略不符。

I0107 11:43:32.129838 14961 layer_factory.hpp:77] Creating layer rpn_loss_cls I0107 11:43:32.132026 14961 net.cpp:150] Setting up rpn_loss_cls I0107 11:43:32.132074 14961 net.cpp:157] Top shape: (1) I0107 11:43:32.132081 14961 net.cpp:160] with loss weight 1 I0107 11:43:32.132100 14961 net.cpp:165] Memory required for data: 298545136 I0107 11:43:32.132104 14961 layer_factory.hpp:77] Creating layer rpn_loss_bbox I0107 11:43:32.132115 14961 net.cpp:106] Creating Layer rpn_loss_bbox I0107 11:43:32.132118 14961 net.cpp:454] rpn_loss_bbox <- rpn_bbox_pred_rpn_bbox_pred_0_split_0 I0107 11:43:32.132123 14961 net.cpp:454] rpn_loss_bbox <- rpn_bbox_targets I0107 11:43:32.132127 14961 net.cpp:454] rpn_loss_bbox <- rpn_bbox_inside_weights I0107 11:43:32.132129 14961 net.cpp:454] rpn_loss_bbox <- rpn_bbox_outside_weights I0107 11:43:32.132133 14961 net.cpp:411] rpn_loss_bbox -> rpn_loss_bbox F0107 11:43:32.132158 14961 smooth_L1_loss_layer.cpp:28] Check failed: bottom[0]->channels() == bottom[1]->channels() (225 vs. 270) *** Check failure stack trace: *** ./experiments/scripts/faster_rcnn_end2end.sh: line 78: 14961 Aborted (core dumped) ./tools/train_net.py --gpu ${GPU_ID} --solver /data/wuxl/RRPN2/models/${PT_DIR}/${NET}/faster_rcnn_end2end/solver.prototxt --weights data/imagenet_models/${NET}.v2.caffemodel --imdb ${TRAIN_IMDB} --iters ${ITERS} --cfg experiments/cfgs/faster_rcnn_end2end.yml ${EXTRA_ARGS}

real 0m18.619s user 0m17.324s sys 0m1.984s

how to solve the problem