• 高效RAW域目标检测挑战赛 - Efficient RAW Object Detection
• • 1. Data Introduction
  • 2. Data Downloading
  • 3. Data Loading
  • • 3.1 for Training Phase
    • 3.2 for Testing Phase

高效RAW域目标检测挑战赛 - Efficient RAW Object Detection

1. Data Introduction

The dataset contains over 25k <image, annotation> pairs from both day and night scenes, and we take approximately 16k as training set, 4k as validation set, and 5k as testing set. Five classes are annotated in the dataset, including Pedestrian, Cyclist, Car, Truck and Tram. The number of objects from each class is listed in the table below.

	Pedestrian	Cyclist	Car	Truck	Tram
Training set	23809	11658	115069	7766	6464
Validation set	2521	1622	24824	2874	1209
Testing set	3752	3569	24246	3476	1249

2. Data Downloading

Downloading from OpenI sometimes is unstable. Thus, we provide alternative downloading links upon Baidu Cloud Disk.

• Training annotations.
• Training raws.
• Validation raws.

3. Data Loading

The entire script for loading data and other utilities can be found in this script.

3.1 for Training Phase

Here is an example of how to load the original raw images (in .raw format) and annotations.

import os
import json
import numpy as np

BIT8, BIT16, BIT24 = 2 ** 8, 2 ** 16, 2 ** 24

def load_raw(name):
    assert os.path.exists(name) and name.endswith('.raw'), \
        f'Invalid raw name < {name} >!'
    
    raw = np.fromfile(name, dtype=np.uint8)
    raw = raw.reshape(1856, 2880, 3).astype(np.float32)
    raw = np.split(raw, 3, axis=2)
    raw = (raw[0] + raw[1] * BIT8 + raw[2] * BIT16) # shape [1856, 2880, 1]
    return raw / (BIT24 - 1) # norm to range [0, 1]


def load_anno(name):
    assert os.path.exists(name) and name.endswith('.json'), \
        f'Invalid annotation name < {name} >!'
    
    with open(name) as f: data = json.load(f)
    for obj in data['shapes']:
        klass = obj['label'] # object class
        xmin, ymin = obj['points'][0] # bounding box
        xmax, ymax = obj['points'][1] # bounding box
        assert 0 <= xmin < xmax <= 2880
        assert 0 <= ymin < ymax <= 1856

3.2 for Testing Phase

Then for submission on our Challenge Homepage, please use the following codes for loading raws. There is a 12MB limit for each image on the testing platform, and the original raws in .raw format is too large (15MB, which is over the limit). Thus, we losslessly compress raws into .npz format. Example codes are as follows.

import os
import numpy as np

BIT8, BIT16, BIT24 = 2 ** 8, 2 ** 16, 2 ** 24

def raw2npz(name):
    assert os.path.exists(name) and name.endswith('.raw'), \
        f'Invalid raw name < {name} >!'
    raw = np.fromfile(name, dtype=np.uint8).reshape(1856, 2880, 3)
    save_name = name.replace('.raw', '.npz')
    np.savez_compressed(save_name, raw)

def load_npz(name):
    assert os.path.exists(name) and name.endswith('.npz'), \
        f'Invalid npz name < {name} >!'
    raw = np.load(name)['arr_0'].astype(np.float32) # shape [1856, 2880, 3]
    raw = np.split(raw, 3, axis=2)
    raw = (raw[0] + raw[1] * BIT8 + raw[2] * BIT16) # shape [1856, 2880, 1]
    return raw / (BIT24 - 1)