RetinaNet_Tensorflow_Rotation icon indicating copy to clipboard operation
RetinaNet_Tensorflow_Rotation copied to clipboard

Focal Loss for Dense Rotation Object Detection

Results 17 RetinaNet_Tensorflow_Rotation issues
Sort by recently updated
recently updated
newest added

在Linux下可以用默认的gcc的编译器编译出box_utils和cython_utils文件夹下的几个文件,但在windows下setup.py文件并不适用,是否有人能提供修改的setup.py文件,使得这个项目可以在windows下运行?

help wanted
good first issue

求助求助! 您好,我使用的裁剪为800×800的DOTA数据集,ResNet50,按照您的教程一步步做的,环境的配置也没有什么问题,但是训练速度大概在2.8秒一张图。这个速度是不是有一些太慢了?想问问各位大佬知道有可能是什么原因吗? 谢谢!!!

here's the equation ![image](https://user-images.githubusercontent.com/28484395/121462520-4c4ff600-c9e3-11eb-83e1-debdf17962dd.png) Let's let `Lreg` as `u` and `|-log(IoU)|` as `I` ![image](https://user-images.githubusercontent.com/28484395/121463142-0cd5d980-c9e4-11eb-8ff6-2e3ccedae27f.png) @yangxue0827 I know there must be something wrong. So could you pleast tell the right wat...

你好,是要先Compile后才能test是么?我在CUDA安装好后进行Compile操作即运行第一个setup.py时报错 OSError: The nvcc binary could not be located in your $PATH. Either add it to your path, or set $CUDAHOME

environments: ubuntu 18.04 cuda 10.0 tensorflow-gpu1.13.1 curexc_type bbox.c:9512:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’; did you mean ‘curexc_value’? tstate->exc_value = local_value; ^~~~~~~~~ curexc_value bbox.c:9513:13: error: ‘PyThreadState...

help wanted

我想利用这个旋转IoU的计算嵌入到我用pytorch写的程序里,请问大佬rbbx_overlaps这个函数的输入有限制吗?

After training the model, how should I calculate the mAP since there is no standard to evaluate besides cls_loss in tensorboard. Hoping to your reply : ) Best regards~

老哥老哥,还有一个疑问,自己想了半天也没明白,网上也找不到: FPN网络里在RoI Pooling层的部分是将不同的RoI分配给金字塔的不同层,使用的分配公式是:`levels = tf.round(4. + tf.log(tf.sqrt(w*h + 1e-8)/224.0) / tf.log(2.))`这里的224是标准ImageNet训练前的图片大小。 所以我的疑问是咱们在训练自己的数据集时,假如我的输入图片大小是800×800,为啥在RoI Pooling的这个公式里还是用的224而不是800呢?这个是跟ResNets用ImageNet预训练出来的有关系吗? 有点懵,望老哥指点一二,抱拳了~

学长你好,请问一下,我用DOTA1.0数据集,配置文件用cfgs_resnet50_dota_v4和cfgs_resnet50_dota_v1,损失函数减小到一定程度后,一直在0.2-1.2震荡。训练50万次后。map如下: This is your result for task 1: mAP: 0.23566566257635685 ap of each class: plane:0.7932061577172662, baseball-diamond:0.1312957338658761, bridge:0.1573459176315261, ground-track-field:0.12121038920876634, small-vehicle:0.40128576370292635, large-vehicle:0.022644009706195248, ship:0.1586747601553599, tennis-court:0.8067936180062999, basketball-court:0.11268627691278782, storage-tank:0.548819002806353, soccer-ball-field:0.0028656588723735567, roundabout:0.07246264775533927, harbor:0.08824166994332554, swimming-pool:0.11577374370476157, helicopter:0.0016795886561954384 请问一下为什么会出现这种情况呢

数据集是HRSC2016 采用配置文件cfgs_res50_dota_v1时loss可以正常收敛,当采用cfgs_res50_dota_v5时cls loss始终在1-1.2之间,reg loss 出现nan,tensorboard中gtboxes_h_gpu、gtboxes_r_gpu_正常有框,positive_anchor也是正常有anchor,但是final_detection_gpu_0始终没有检查框,请问是什么问题?非常感谢!