YOLOX中decode 特征点解码过程可视化

weixin_45734021

已于 2024-10-10 22:23:47 修改

阅读量559

点赞数 6

文章标签： python numpy 开发语言 YOLO

于 2024-10-10 21:52:28 首次发布

本文链接：https://blog.youkuaiyun.com/weixin_45734021/article/details/142833036

版权

YOLOX中decode 特征点解码过程可视化
该代码是特征宽高为20*20，batch_size=4，num_classes = 20进行解码可视化的过程。

import numpy as np
import matplotlib.pyplot as plt

def decode_for_vision(output):
    bs, hw = np.shape(output)[0], np.shape(output)[1:3]
    #  hw[0] * hw[1]  ------- 20,20
    output = np.reshape(output, [bs, hw[0] * hw[1], -1])
    #print(output)
    #output ------(4, 400, 23)
    grid_x, grid_y = np.meshgrid(np.arange(hw[1]), np.arange(hw[0]))
    #print(grid_x)
    grid = np.reshape(np.stack((grid_x, grid_y), 2), (1, -1, 2))
    #grid ---------(1, 400, 2)
    #print(grid)
    box_xy = (output[..., :2] + grid)
    #box_xy.shape      (4, 400, 2)
    #output[..., :2]    (4, 400, 2)
    #grid               (1, 400, 2)
    box_wh = np.exp(output[..., 2:4])
    #output[..., 2:4].shape       (4, 400, 2)
    #box_wh                       (4, 400, 2)
    fig = plt.figure()
    ax = fig.add_subplot(121)
    plt.ylim(-2.22, hw[0] + 2.22)
    plt.xlim(-2.22, hw[1] + 2.22)
    plt.scatter(grid_x, grid_y)
    plt.scatter(0, 0, c='black')
    plt.scatter(1, 0, c='black')
    plt.scatter(2, 0, c='black')
    plt.scatter(box_xy[0, 0, 0], box_xy[0, 0, 1], c='r')
    plt.scatter(box_xy[0, 1, 0], box_xy[0, 1, 1], c='g')
    plt.scatter(box_xy[0, 2, 0], box_xy[