DOTA v1.0 test results achieved 0.620, a little far from 0.723 about bbavectors-oriented-object-detection HOT 17 CLOSED

是这个网址，服务器不是很稳定
可能batchsize比原文大，效果会相对好一点。

• mAP: 0.7386079289654184
• ap of each class: plane:0.8906786852353635, baseball-diamond:0.7973180557325118, bridge:0.49555937732787664, ground-track-field:0.6652836232472856, small-vehicle:0.7719567850263218, large-vehicle:0.8127366944445986, ship:0.8785596536225615, tennis-court:0.9087453350749631, basketball-court:0.86021476336708, storage-tank:0.8668640701621114, soccer-ball-field:0.5594583780845677, roundabout:0.6310733396356176, harbor:0.6609224804633762, swimming-pool:0.6937428143102882, helicopter:0.5860048787467512

from bbavectors-oriented-object-detection.

54717张图像是不是忘记了0.5 的尺度了？
split.splitdata(1)
加一个
split.splitdata(0.5)

from bbavectors-oriented-object-detection.

54717张图像是不是忘记了0.5 的尺度了？
split.splitdata(1)
加一个
split.splitdata(0.5)

加上之后舍去不包含目标的图像共69337是吗？请问你复现之后是否能达到73的mAP呢

from bbavectors-oriented-object-detection.

没舍弃共69377张，刚开始训 还没出结果

from bbavectors-oriented-object-detection.

没舍弃共69377张，刚开始训 还没出结果

谢谢，刚加上0.5的倍率之后数目一致了

from bbavectors-oriented-object-detection.

Hello,
I have trained the model followed by your settings with a batch size of 20 on 4 RTX 2080Ti GPUs. I cropped the images into 600x600 with a stride of 100 using ImgSpit_multi_process.py of the official repo, but only got 54717 images on trainval when preserving negative samples with no objects in them. Finally after 40 epochs of training, I got mAP of 61.95 on test dataset, a little far from 72.32 showed in the paper. Can you help me with that? Besides, I can't got 69377 cropped images of trainval set. Thank you.

请问怎么测试的？ https://captain-whu.github.io/DOTA/evaluation.html 这个网址点不开咋回事？

from bbavectors-oriented-object-detection.

是这个网址，服务器不是很稳定

from bbavectors-oriented-object-detection.

是这个网址，服务器不是很稳定
可能batchsize比原文大，效果会相对好一点。

• mAP: 0.7386079289654184
• ap of each class: plane:0.8906786852353635, baseball-diamond:0.7973180557325118, bridge:0.49555937732787664, ground-track-field:0.6652836232472856, small-vehicle:0.7719567850263218, large-vehicle:0.8127366944445986, ship:0.8785596536225615, tennis-court:0.9087453350749631, basketball-court:0.86021476336708, storage-tank:0.8668640701621114, soccer-ball-field:0.5594583780845677, roundabout:0.6310733396356176, harbor:0.6609224804633762, swimming-pool:0.6937428143102882, helicopter:0.5860048787467512

It's amazing. Can you share the training batch size?

from bbavectors-oriented-object-detection.

是这个网址，服务器不是很稳定
可能batchsize比原文大，效果会相对好一点。

• mAP: 0.7386079289654184
• ap of each class: plane:0.8906786852353635, baseball-diamond:0.7973180557325118, bridge:0.49555937732787664, ground-track-field:0.6652836232472856, small-vehicle:0.7719567850263218, large-vehicle:0.8127366944445986, ship:0.8785596536225615, tennis-court:0.9087453350749631, basketball-court:0.86021476336708, storage-tank:0.8668640701621114, soccer-ball-field:0.5594583780845677, roundabout:0.6310733396356176, harbor:0.6609224804633762, swimming-pool:0.6937428143102882, helicopter:0.5860048787467512

最后我的mAP是69.97，40个epoch时loss还有下降空间，我再继续训练看看

from bbavectors-oriented-object-detection.

和paper中的结果相比，有的类别有上升而有的类别有较大下降

from bbavectors-oriented-object-detection.

和paper中的结果相比，有的类别有上升而有的类别有较大下降

好的，你再试一下，我上面的结果没有改什么，你有在HRSC的数据集上run吗? 我跑的结果 epoch=100(只用训练集map87.6/88.15两次结果，训练集+测试集map89.14)。感觉波动性挺大的，作者的意思应该是只用了训练集(paper,88.6)

from bbavectors-oriented-object-detection.

和paper中的结果相比，有的类别有上升而有的类别有较大下降

你可以把 --conf_thresh调低一点，0.1或者0.05

from bbavectors-oriented-object-detection.

和paper中的结果相比，有的类别有上升而有的类别有较大下降

你可以把 --conf_thresh调低一点，0.1或者0.05

You are right. After I changed the conf_thresh from 0.18 to 0.1, the mAP increased to 71.32.

from bbavectors-oriented-object-detection.

和paper中的结果相比，有的类别有上升而有的类别有较大下降

好的，你再试一下，我上面的结果没有改什么，你有在HRSC的数据集上run吗? 我跑的结果 epoch=100(只用训练集map87.6/88.15两次结果，训练集+测试集map89.14)。感觉波动性挺大的，作者的意思应该是只用了训练集(paper,88.6)

HRSC还没跑，感觉跟数据关系挺大的

from bbavectors-oriented-object-detection.

和paper中的结果相比，有的类别有上升而有的类别有较大下降

你可以把 --conf_thresh调低一点，0.1或者0.05

You are right. After I changed the conf_thresh from 0.18 to 0.1, the mAP increased to 71.32.

You can also increase the batch_size, I got mAP of 74.92 after using batch_size = 48 on 4 RTX6000.

from bbavectors-oriented-object-detection.

@zhangyuxuan1996 我在训练hrsc后，测试的可视化效果里边界框都是平行四边形，而不是像gt那样的旋转矩形框，是怎么回事呢，请问你在训练后测试可视化有没有遇到同样的情况呢

from bbavectors-oriented-object-detection.

I get the mAP 69.01 after using batch_size = 64 on 8 V100. the lr_rate is 2e-4, I trained it with 60 epoch and using the learning schedule as shown below:

        self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, mode='min', factor=0.5, patience=5)

The conf_thresh is 0.18

from bbavectors-oriented-object-detection.