python implementation

[HuaiyuanXu]

Could you provide a python implementation for preprocessing?

[VenkateshSoni]

Can you please elaborate the issue more ?

[HuaiyuanXu]

Many thanks for your reply. Could you provide code in python language to generate global descriptors? In other words, could you give the python language code to replace the c++ language code in the gen_desc folder?

[lilin-hitcrt]

Could you provide a python implementation for preprocessing?

Sorry for not replying in time as I was preparing for ICRA2023 some time ago. I will implement and upload the relevant code as soon as possible.

[lilin-hitcrt]

Could you provide a python implementation for preprocessing?

I'll complete the Python implementation for preprocessing at my earliest convenience.

[lilin-hitcrt]

Perhaps this code will work.

import numpy as np
import yaml

class PreProcessor:
    def __init__(self, cfg_file):
        with open(cfg_file, 'r') as file:
            cfg = yaml.safe_load(file)
        self.learning_map = None
        if cfg['remap']:
            cfg_lm = cfg['learning_map']
            max_id = np.max(list(cfg_lm.keys()))
            learning_map_list = [cfg_lm[i] if i in cfg_lm else -1 for i in range(max_id + 1)]
            self.learning_map = np.array(learning_map_list, dtype='int')
        self.valid_labels = np.array([1] + list(range(9, 20)), dtype='int')
        self.label_map = np.full(20, -1, dtype='int')
        self.label_map[9:] = np.arange(1, 12)
        self.label_map[1] = 0

    def run(self, pcd_file, label_file):
        # load data
        points = self.load_pcd(pcd_file)
        labels = self.load_label(label_file)

        # remove unused label
        label_mask = np.isin(labels, self.valid_labels)
        points = points[label_mask]
        labels = labels[label_mask]

        # Remove distant data
        dis = np.linalg.norm(points[:, :2], axis=1)
        dis_mask = (dis <= 50)
        points = points[dis_mask]
        labels = labels[dis_mask]
        dis = dis[dis_mask]

        angles = np.arctan2(points[:, 1], points[:, 0]) + np.pi
        angles = (angles * 180.) / np.pi
        angles = np.floor(angles).astype('int')

        #Ensure the angle is valid
        angle_mask = ((angles >=0) & (angles <= 359))
        points = points[angle_mask]
        labels = labels[angle_mask]
        dis = dis[angle_mask]
        angles = angles[angle_mask]

        # convert label id to 0-11
        labels = self.label_map[labels]
        order = np.argsort(-dis)
        dis = dis[order]
        points = points[order]
        labels = labels[order]
        angles = angles[order]
        # desc for semantics other than road
        min_desc = np.zeros([12, 360], dtype='float')
        min_desc[labels, angles] = dis
        order = np.argsort(dis)
        dis = dis[order]
        points = points[order]
        labels = labels[order]
        angles = angles[order]
        # desc for road
        max_desc = np.zeros([12, 360], dtype='float')
        max_desc[labels, angles] = dis
        min_desc[1] = max_desc[1] # road
        return min_desc

    def load_pcd(self, file):
        points = np.fromfile(file, dtype='float32').reshape(-1, 4)[:, 0:3]
        return points

    def load_label(self, file):
        labels = np.fromfile(file, dtype='uint32')
        if self.learning_map is not None:
            labels = labels & np.uint32(0x0000FFFF)
            labels = self.learning_map[labels]
        return labels

if __name__ == '__main__':
    processor = PreProcessor("gen_desc/conf/sem_config.yaml")
    desc = processor.run('/home/l/workspace/SSC/data/000720.bin', '/home/l/workspace/SSC/data/000720.label')
    from matplotlib import pyplot as plt
    plt.imshow(desc)
    plt.show()