Skip to content

Object detection, 3D detection, and pose estimation using center point detection:

License

Notifications You must be signed in to change notification settings

sisrfeng/CenterNet

 
 

Repository files navigation

生成高斯热图

/home/wf/d/clwf/src/lib/utils/image.py

image

CenterPoint的作者在issue里回复说, 影响不大. 他们设为固定的大小, 比如2,

image


1

my batch_size is 8.Does it matter?
It matters a lot.

2

https://blog.csdn.net/weixin_42634342/article/details/97756458

  1. opt.py应该是opts.py
  2. “由于图片路径问题,要在test函数里面修改一下数据集图片的绝对路径”,没找到相应代码。 3.不是修改,而是自己加一行dataset.img_dir='/d/COCO_my'

上面的博客不详细,参考这个https://blog.csdn.net/weixin_41765699/article/details/100118353:
一开始被 https://img-blog.csdnimg.cn/20190828210637735.png 误导了 其实,仿照coco.py自己建立的aic.py文件中相应部分应该是

valid_classids
        self.class_name = [
            '__background__', 'car']
        self._valid_ids = [
            3]

coco中, car对应类别3,aic中一致。如果把对应背景的0放进_valid_ids,训练时报错。

3

default_dataset_info = {
对应原repo的

      'ctdet': {'default_resolution': [512, 512], 'num_classes': 80, 
                'mean': [0.408, 0.447, 0.470], 'std': [0.289, 0.274, 0.278],
                'dataset': 'coco'},

运行demo.py时要在opts.py中改回来,否则提示类别数目不对。

4

根据这个转数据集 /src/tools/convert_kitti_to_coco.py。

5

学习率:
The learning rate 5e-4 is set for batchsize 128, you will need to linearly scale it to your batchsize.

6

https://blog.csdn.net/weixin_41765699/article/details/100118353中, 指定--master_batch 1
, 这样batch就不会平均分配到各个GPU。master GPU要存model的参数,平均分配batchsize不合理。

7

如果采用了multi-scale测试,需要使用softnms作为后处理。要额外编译

8

https://blog.csdn.net/weixin_42634342/article/details/97756458
说运行test.py时,不要加上 --keep_res。
但用作者给的模型以及自己训的模型,在指定--input_res 512和--keep_res测试时,测出来的AP都正常。感觉这篇博客不太可靠。

9

1.关于ctdet_coco_dla_1x.sh和ctdet_coco_dla_2x.sh中提到的训练设置:
1x是训练了140个epoch,在90和120个epoch的时候学习率变为原来的十分之一。
2x的训练,有两种: 方法一,是直接训练230个epoch,是在180和210个epoch的时候进行学习率衰减; 方法二,是在1x训练的基础上进行finetune,即继续原来的训练;

10 关于finetune:

I haven't tried freezing any layers.
You can load the COCO pretrained model by --load_model ../models/ctdet_coco_2x.pth when finetuning on your own dataset.

从头训练的学习率: should set lr to 1.25e-2 when batch size is 32,
and scale up/down linearly according to this reference numbers.
xingyizhou#315 (comment)

11

xingyizhou#660 We WILL predict peaks with very low confidence. This is not punished at all for the AP metric. However I do have tried to set a 0.01 threshold, and the results stayed the same for the specific model I tested.

About

Object detection, 3D detection, and pose estimation using center point detection:

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 77.1%
  • Cuda 7.7%
  • C++ 6.9%
  • C 6.6%
  • Shell 1.0%
  • Cython 0.7%