README.md

metadata

extra_gated_prompt: Test
extra_gated_fields:
  Name: text
  PI/Advisor: text
  Affiliation: text
  Purpose: text
  Country: country
  I agree to use this dataset for non-commercial use ONLY: checkbox

---

Dataset Card for Dataset CrashCar

This is the dataset proposed in 'CrashCar101: Procedural Generation for Damage Assessment' [WACV24]

• Project Page: https://crashcar.compute.dtu.dk
• Repository: https://github.com/JensPars/CrashCar_procedural_generation
• Paper: https://openaccess.thecvf.com/content/WACV2024/papers/Parslov_CrashCar101_Procedural_Generation_for_Damage_Assessment_WACV_2024_paper.pdf

Example dataset class in pytorch

import os
import torch
from glob import glob
from PIL import Image
import numpy as np
from pathlib import Path
import pandas as pd

class CarDataset(torch.utils.data.Dataset):
    def __init__(self, root_dir, transform=None, tgt_transform=None):
        img_root = os.path.join(root_dir, 'img', '*', '*.png')
        part_root = os.path.join(root_dir, 'parts', '*', '*.png')
        damage_root = os.path.join(root_dir, 'damage', '*', '*.png')
        self.img_root = sorted(glob(img_root))
        self.part_root = sorted(glob(part_root))
        self.damage_root = sorted(glob(damage_root))
        self.transform = transform
        self.tgt_transform = tgt_transform

    def __len__(self):
        return len(self.img_root)

    def __getitem__(self, idx):
        img = Image.open(self.img_root[idx])
        part_img = Image.open(self.part_root[idx])
        damage_img = Image.open(self.damage_root[idx])
        
        if self.transform:
            img = self.transform(img)
            part_img = self.transform(part_img)
            damage_img = self.transform(damage_img)

        return {
            'image': img,
            'part': part_img,
            'damage': damage_img
        }

The following code will yield

import matplotlib.pyplot as plt
import numpy as np

dataset = CarDataset(root, transform=None)
out = dataset[20000]

fig, axs = plt.subplots(1, 3, figsize=(15, 5))

axs[0].imshow(out['image'])
axs[0].axis('off')

axs[1].imshow(out['image'])
alpha_map = (np.array(out['damage'])!= 0).astype(float)
axs[1].imshow(out['damage'], cmap="jet", alpha=alpha_map)
axs[1].axis('off')

axs[2].imshow(out['image'])
alpha_map = (np.array(out['part'])!= 0).astype(float)
axs[2].imshow(out['part'], cmap="jet", alpha=alpha_map)
axs[2].axis('off')

plt.tight_layout()
plt.show()