Dataset Viewer
Auto-converted to Parquet Duplicate
image
imagewidth (px)
1.02k
1.02k
source_path
stringclasses
8 values
frame_id
int64
100
107
drive
stringclasses
1 value
split
stringclasses
1 value
DurLAR_20210716_S/image_01/data/0000000100.png
100
DurLAR_20210716_S
image_01
DurLAR_20210716_S/image_01/data/0000000101.png
101
DurLAR_20210716_S
image_01
DurLAR_20210716_S/image_01/data/0000000102.png
102
DurLAR_20210716_S
image_01
DurLAR_20210716_S/image_01/data/0000000103.png
103
DurLAR_20210716_S
image_01
DurLAR_20210716_S/image_01/data/0000000104.png
104
DurLAR_20210716_S
image_01
DurLAR_20210716_S/image_01/data/0000000105.png
105
DurLAR_20210716_S
image_01
DurLAR_20210716_S/image_01/data/0000000106.png
106
DurLAR_20210716_S
image_01
DurLAR_20210716_S/image_01/data/0000000107.png
107
DurLAR_20210716_S
image_01

DurLAR_S in Lance

This release converts the public DurLAR_S snapshot to Lance. It includes 3,000 frames across five drives with left and right camera imagery, panoramic ambient and reflectivity imagery, and LiDAR point clouds. The original GPS, IMU, lux, and timestamp files are available under source/.

Data layout

Each drive has 600 rows in each of its five Lance tables under data/<drive>/. The four image tables retain the original Hugging Face split names. The additional lidar.lance table retains the original point cloud files.

Table Original files
image_01.lance Right camera PNG
image_02.lance Left RGB camera PNG
ambient.lance Panoramic near-infrared PNG
reflectivity.lance Panoramic reflectivity PNG
lidar.lance KITTI-compatible point cloud BIN

The same frame_id identifies matching frame filenames within a drive. Use each modality's timestamps for time-based synchronization. The original column stores the original binary file. Each row records its upstream path, size, and SHA-256 digest. The dataset viewer displays the first eight original images from each image split in Parquet, and the Lance tables hold the complete snapshot.

Lance stores 3,000 original files directly in its Blob v2 storage. Blob v2 references serve the other 12,000 original files from verified copies under source/ in this same repository.

Search for similar road scenes

The 3,000 left RGB frames have generated 1,152-dimensional, L2-normalized float32 embeddings from the image encoder of Google SigLIP 2 So400m, patch16-384, pinned to revision dd658faac399427308559e2c3ac1e99cbe43845d. Images are decoded to RGB and resized by the model's processor. Search uses cosine distance, with similarity reported as one minus distance. Matching frame and drive identifiers also locate the point clouds and other image modalities. Pinned embedding model

Install the required Python packages with pip install 'pylance>=14.0.3' torch transformers pillow numpy. This example embeds a camera frame from the published dataset and returns the five closest other frames across all drives.

import io

import lance
import torch
from PIL import Image
from transformers import AutoModel, AutoProcessor

root = "hf://datasets/lance-format/durlar-s-lance"
drives = ("20210716", "20210901", "20211012", "20211208", "20211209")
tables = [lance.dataset(f"{root}/data/DurLAR_{drive}_S/image_02.lance") for drive in drives]
source = tables[0]
query_index = 100
query_path = source.take([query_index], columns=["source_path"])["source_path"][0].as_py()
_, payload = source.read_blobs("original", indices=[query_index])[0]
with Image.open(io.BytesIO(payload)) as decoded:
    image = decoded.convert("RGB")

model_id = "google/siglip2-so400m-patch16-384"
revision = "dd658faac399427308559e2c3ac1e99cbe43845d"
processor = AutoProcessor.from_pretrained(model_id, revision=revision)
model = AutoModel.from_pretrained(model_id, revision=revision).eval()
with torch.inference_mode():
    features = model.get_image_features(**processor(images=[image], return_tensors="pt")).pooler_output
    query = torch.nn.functional.normalize(features.float(), dim=-1)[0].numpy()

results = []
for table in tables:
    rows = table.to_table(
        columns=["drive", "frame_id", "source_path", "_distance"],
        nearest={"column": "embedding", "q": query, "k": 6, "metric": "cosine"},
    ).to_pylist()
    results.extend(row for row in rows if row["source_path"] != query_path)
for row in sorted(results, key=lambda item: item["_distance"])[:5]:
    print(f"{1 - row['_distance']:.4f}\t{row['drive']}\tframe {row['frame_id']}\t{row['source_path']}")

Source and license

DurLAR was created by Li Li, Khalid N. Ismail, Hubert P. H. Shum, and Toby P. Breckon at Durham University. The original publication appeared in the Proceedings of the International Conference on 3D Vision, 2021. Original media and sensor files are shared under Creative Commons Attribution 4.0. This conversion groups original payloads into Lance files and adds search vectors and viewer samples. Original Hugging Face dataset · Pinned source revision · University collection · License and attribution

Downloads last month
-