这是我第一次参加天池的大赛，半决赛的代码开源在了final_commit文件夹里面包含了填鸭的代码，第一版的填鸭，计算了patch块的相似度。第二版的我们对小目标（1-4）类的随机3-5倍的放大。 1.半决赛的Rank：34/100 2.线下的map：60%左右，线上的map：40%左右
This is the first time I participated in the Tianchi competition. The semi-final code is open sourced. The final_commit folder contains the code for duck filling. In the second version, we zoomed in randomly 3-5 times on the small target (1-4).
Rank of semi-final: 34/100
Offline Map: about 60%, online map: about 40%
-
特征工程
-
选用模型
-
训练，调参
-
提交结果
开始做的时候，我们是先出了一个baseline的结果，开始我是自己一个人玩，直接上faster-rcnn-r101 map：26% 有点沮丧 毕竟这时候还在忙着写（水）论文，后来论坛里 开源了一个cascade-rcnn-r50的模型。初赛52%map。 根据这个baseline，我换了backbone r101 居然：54%map，嘻嘻...这里直接就进了60多。
Feature engineering
-
** Selection model **
-
** Training and Tuning **
-
** Submission Results **
**When we started, we first produced a baseline result. At the beginning, I played by myself, and went directly to the faster-rcnn-r101 map: 26% a bit frustrated ** **After all, at this time, I was still busy writing a (water) dissertation. Later, a cascade-rcnn-r50 model was open sourced in the forum. 52% map for the preliminary round. ** According to this baseline, I changed the backbone r101: 54% map, ... more than 60 directly entered here .
-
- anchor的设计非常重要，mmdetection的默认[0.5,1,2]一般来说很难符合数据的特性，所以这里是提分的点子
- fpn层 dcn （槽点，太吃显存，因为要用很多的offsets）
- OHEM 在线困难样本的发掘
- soft-nms 提分不多，大概一个点左右（大概率是0.几%哈~）
- TTA 老版本的mmdet没有TTA 多尺度测试，新版的有
- 填鸭，对于正常样本的利用。这里其实跟我写论文里的东西有点像，来自小样本增强的那篇论文，但是有个问题就是容易引入结构化的噪声，这里需要计算 patch的块的相似度---于是乎（度娘了一下），用了现成的
- GN+ws （分组Normalize）
- rpn 调参
-The design of the anchor is very important. The default [mm, detection] of mmdetection is generally difficult to meet the characteristics of the data, so here is a point to improve
fpn layer dcn (slot, too much video memory, because a lot of offsets are used)
Discovery of OHEM online difficult samples
Soft-nms does not improve much, about one point (large probability is 0. several% ha ~)
-TTA old version of mmdet does not have TTA multi-scale test, new version
Duck filling, use of normal samples. This is actually a bit like what I wrote in the paper, from the paper with small sample enhancement, but there is a problem that it is easy to introduce structured noise. Here you need to calculate the similarity of the patch blocks-so it ’s (Baidu (google) A little bit), using ready-made
GN + ws (group Normalize)
rpn tuning
-
打算切图做的，cascade-r50对512*512的图像来测试，对于1-4好像有点效果，可能是可视化的错觉吧
-
分类的模型，这时候就做的比较晚，用了 20层的 只有40多的Acc
-I plan to cut the picture. Cascade-r50 tests on 512 * 512 images. It seems to be a bit effective for 1-4 classes. It may be a visual illusion.
-The classification model, which was done relatively late at this time, used 20 layers and only had more than 40 Acc+
- 其实需要魔改框架的，加入一些对小目标增强的结构
- ** In fact, it is necessary to magically change the framework, and add some structures that enhance small goals **
- 参考上传的 transform.py 替换 mmdet的同名文件
- ** Refer to uploaded transform.py to replace mmdet's file with the same name **
-
Install
Fellow the mmdetection install.md
-
Data format
In my experiment COCO, VOC changed by yourself!
-
COCO pretrained model transfer
The transfer code in checkpoints num_class should modify your class. Notice!!! scale and ratios changed, I use a simple cat method for suitable model struct param avoid parameters initialize problem!
python3 configs/cascade_rcnn_r101_fpn_1x_with_coco.py --gpus 1 work_dir XXXXXX(your path to save model and train log)python3 cascade_rcnn_r101_fpn_1x_test_coco.py
Thanks TianChi for holding this competition https://tianchi.aliyun.com/
Thanks Openbyes provide computing power support https://openbayes.com/
Thanks Team member Chen and Li give valuable advice
Email：[email protected]