I noted a comprehensive autonomous‑driving benchmark that highlights the gap between lab results and real‑world performance across several perception tasks.
Dataset size ≈ 3 TB.
Contains 389 stereo and optical‑flow image pairs.
Provides stereo visual‑odometry sequences covering 39.2 km of driving.
Includes over 200 k 3D object annotations captured in cluttered urban scenarios.
Evaluations showed that state‑of‑the‑art algorithms failed when they moved from lab to the real world, exposing inherent bias in current methods.
SOTA algorithms failed when they moved from lab to the real world.
Camera type - fisheye, 4 cameras covering 360 degree covering wide field of view
Free of charge for public use
Focused on tasks such as 2D/3D object detection, semantic segmentation and depth estimation
It is complementary to datasets such as KITTI, Cityscapes, Mapillary, Apolloscape, BDD100k, nuScenes etc especially because cameras used here have narrow field of view!
I announced the Roadside Perception 3D (RoPe3D) dataset on 2022‑10‑26 as a fresh autonomous‑driving benchmark that tackles the blind spots of traditional frontal‑view camera setups.
RoPe3D.full_form = Roadside Perception 3D dataset
RoPe3D.name = RoPe3D
existing_datasets.camera_placement = frontal view
existing_datasets.limitations = "cannot observe surroundings for a long period due to movement; perceptual range is limited and vulnerable to occlusion due to low height"
RoPe3D.camera_placement = roadside pole‑mounted cameras a few meters above ground
RoPe3D.image_count = 50k images
RoPe3D.object_3d_count = over 1.5M 3D objects
I decided to place cameras on roadside poles a few metres high to obtain a continuous, wide‑angle view with minimal occlusion, directly addressing the limitations of low‑height frontal mounts.