Zhao-Yian

Update predict_system.py

3

修改思路在issue101中详细描述

主体检测推理代码存在bug

1

代码位置：PaddleClas/deploy/python/predict_system.py 94-98行： ` # st1: get all detection results results = self.det_predictor.predict(img) # st2: add the whole image for recognition to improve recall results = self.append_self(results, img.shape) ` 这里，首先将主体检测的检测结果放入results，然后为了提高精度，又添加了一张整图的shape。这样理论上没有问题，如果主体检测推理的bbox都不合适的话，整张图像送入特征提取模型一定会得到一个比较好的结果，因为索引库的编码都是整图编码。但是经过多次实验发现，主体检测大多数情况下性能都不错，这样的话检测到效果最好的图像主体分数一般接近1，但是整图的分数会更高，之后nms会将最好的bbox过滤掉，最后输出的可视化结果如下图所示： ...

Evaluation scripts

8

Can evaluation scripts be provided on different datasets to validate the quantitative results provided in the paper？

code

1

I am glad to hear that your paper has been accepted. When will you make the code public?

This is a very good work! I used the official weights to export onnx, and the specified version of tensorrt export engine reported an Segmentation fault (core dumped) error. The...