'BoxList' object has no attribute 'quad_bbox'

Xiangyu-CAS / R2CNN.pytorch

pytorch implementation of R2CNN, Rotational Faster RCNN for orientated object detection

MIT License

61 stars 10 forks source link

'BoxList' object has no attribute 'quad_bbox' #2

Open yoyoyo-yo opened 4 years ago

yoyoyo-yo commented 4 years ago

❓ Questions and Help

yoyoyo-yo commented 4 years ago

Thank you for your gread job!

I installed and used inference_engine.py. Then, I got

Traceback (most recent call last):
  File "inference_engine.py", line 230, in <module>
    canvas = detector.run_on_opencv_image(img)
  File "inference_engine.py", line 93, in run_on_opencv_image
    result = self.overlay_boxes(result, top_predictions)
  File "inference_engine.py", line 172, in overlay_boxes
    quad_boxes = predictions.quad_bbox
AttributeError: 'BoxList' object has no attribute 'quad_bbox'

What shold I do?

AlfengYuan commented 4 years ago

This implement doesn't include 11x3 and 3x11 RoiAligan or RoiPooling, and inclined nms agorithm. This is a good project for learning, but it is too far from the original paper.

Xiangyu-CAS commented 4 years ago

@yoyoyo-yo It seems like you have installed the original version of MaskRCNN-benchmark, so it points to another BoxList class, which has no implementation of Quad_bbox.

Xiangyu-CAS commented 4 years ago

@yyfhaku This version only implemented the major though of R2CNN, that's the rotated region regression. Actually, there are still lots of hyperparamters and tricks need to be carefully adjusted.

BTW: 113 and 311 proved to be not working well in general case

yangkai798 commented 4 years ago

嗨，请问这里最后预测的还是按照8个点来吗？而不是论文中说的, x1, y1, x2, y2 和h? Inclined NMS 算IoU的时候实际上就是Skew IoU 是吗？多谢！

AlfengYuan commented 4 years ago

是啊，你是对的，这个跟论文完全不一样，但是可以学习，可以参考它的自己定义数据流，试着去复现

---Original--- From: "yangkai798"<notifications@github.com> Date: Mon, Dec 30, 2019 16:34 PM To: "Xiangyu-CAS/R2CNN.pytorch"<R2CNN.pytorch@noreply.github.com>; Cc: "Mention"<mention@noreply.github.com>;"yyfhaku"<2640706467@qq.com>; Subject: Re: [Xiangyu-CAS/R2CNN.pytorch] 'BoxList' object has no attribute 'quad_bbox' (#2)

嗨，请问这里最后预测的还是按照8个点来吗？而不是论文中说的, x1, y1, x2, y2 和h? Inclined NMS 算IoU的时候实际上就是Skew IoU 是吗？多谢！

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

yoyoyo-yo commented 4 years ago

@Xiangyu-CAS I followed your INSTALL.md, but got same error. This looks like maskrcnn_benchmark path error?

xiabc612 commented 4 years ago

There is no need to clone another maskrcnn-benchmark, just run "python setup.py build develop" in root dir. The installation instructions need to be improved!

Xiangyu-CAS commented 4 years ago

@xiabc612 you are right... I didn't update installation.md, so it's the same as maskrcnn-benchmark. I will revise it. Thank you

Xiangyu-CAS commented 4 years ago

@yoyoyo-yo as xiabc612 said， there's no need to git clone another maskrcnn-benchmark, I have updated installation instructions. It's my fault, sorry for it.

Xiangyu-CAS commented 4 years ago

@yangkai798 是按照4个点8个参数来的，[x1, y1, x2, y2, h]也可以，后者效果在ICDAR上更好一点点，但区别应该不大，用在遥感里大家更倾向用4个点。 Inclinded NMS这里还没实现...

YQ-Yang commented 3 years ago

请问这个能否在自己的数据集实现？