MSherbinii/ipad-industrial-anomaly

IPAD: Industrial Process Anomaly Detection Dataset

This is the official IPAD dataset for video anomaly detection in industrial scenarios.

Key Features

• 16 Industrial Devices: 12 synthetic + 4 real-world scenarios
• 597,979 Video Frames: Over 6 hours of footage
• Periodicity Annotations: Unique temporal features for industrial processes
• 39 Anomaly Classes: Various types including:
  • Appearance anomalies (color, shape changes)
  • Position anomalies (shifts, tilts)
  • Motion anomalies (lagging, speed changes)
  • Logic anomalies (sorting errors, wrong sequences)

Dataset Structure

ipad_dataset/
├── S01_conveyer_belt/
├── S02_automatic_lifter/
├── S03_forklift_truck/
├── S04_manual_cutter/
├── S05_90_conveyer/
├── S06_180_conveyer/
├── S07_z_lifter/
├── S08_box_sorter/
├── S09_mechanical_gripper/
├── S10_standing_crane/
├── S11_automatic_cutter/
├── S12_drilling_machine/
├── R01_conveyer_belt_real/
├── R02_automatic_lifter_real/
├── R03_forklift_truck_real/
└── R04_manual_cutter_real/

Each device folder contains:

• training/frames/: Normal operation videos
• testing/frames/: Both normal and anomalous videos
• Frame-level annotations for anomalies

Statistics

• Total Frames: 597,979
• Training Frames: 430,867 (normal only)
• Testing Frames: 167,112 (normal + anomalies)
• Resolution: 2492 × 988 pixels
• Frame Rate: Variable (5-30 second cycles)

Differences from Existing Datasets

Unlike human-centric VAD datasets (UCSD Ped, Avenue, ShanghaiTech):

• ✅ Object-centric: Anomalies can occur anywhere in frame
• ✅ Periodic: Equipment has cyclic motion patterns
• ✅ Industrial-specific: Real factory scenarios
• ✅ Synthetic + Real: Domain adaptation capability

Usage

Download and Extract

from huggingface_hub import hf_hub_download
import zipfile

# Download the dataset
zip_path = hf_hub_download(
    repo_id="MSherbinii/ipad-industrial-anomaly",
    filename="ipad_dataset.zip",
    repo_type="dataset"
)

# Extract
with zipfile.ZipFile(zip_path, 'r') as zip_ref:
    zip_ref.extractall("./ipad_data")

Load with PyTorch

from torch.utils.data import Dataset
import cv2
import os

class IPADDataset(Dataset):
    def __init__(self, root_dir, device="S01", split="train"):
        self.root = os.path.join(root_dir, f"{device}/{split}/frames")
        # ... implementation

    def __getitem__(self, idx):
        # Load 16-frame video clip
        frames = []
        for i in range(idx, idx+16):
            frame = cv2.imread(f"{self.root}/frame_{i:05d}.jpg")
            frame = cv2.resize(frame, (256, 256))
            frames.append(frame)
        return torch.tensor(frames)

Baseline Results

From the original paper using Video Swin Transformer + Periodic Memory:

Device	AUC (%)	Device	AUC (%)
S01	69.5	S07	60.6
S02	63.9	S08	85.6
S03	70.6	S09	71.2
S04	58.3	S10	62.2
S05	86.2	S11	60.9
S06	61.2	S12	67.1
R01	84.4	R03	43.5
R02	75.4	R04	76.7
Average	68.6

Citation

@article{liu2024ipad,
  author    = {Jinfan Liu, Yichao Yan, Junjie Li, Weiming Zhao,
               Pengzhi Chu, Xingdong Sheng, Yunhui Liu and Xiaokang Yang},
  title     = {IPAD: Industrial Process Anomaly Detection Dataset},
  year      = {2024},
  journal   = {arXiv preprint arXiv:2404.15033},
}

License

Please refer to the original paper and project page for license information.

Acknowledgments

• Original dataset creators: Shanghai Jiao Tong University & Lenovo Research
• Uploaded to HuggingFace Hub by: MSherbinii
• Part of the IPAD VAD reproduction and improvement project