prakashchhipa/ImageNet-PD

ImageNet-PD is a new benchmark dataset for robustness evaluation against unprecedented changes in shape, size, orientation, angles, and other spatial relationships of visual concepts in images due to perspective distortion. ImageNet-PD is dedicated benchmark to evaluate robustness of models for perspective distortion, derived from the ImageNet validation set by synthesizing distortions in different orientations . ImageNet-PD has eight subsets four corresponding to four orientations (left, right, top, bottom) with black background (PD-L, PD-R, PD-T, PD-B). The other four subsets are with the same orientations but with integrated padding background using boundary pixels (PD-LI, PD-RI, PD-TI, PD-BI).

Why ImageNet-PD? Perspective distortion (PD) is pervasive in real-world imagery and presents significant challenges for developing computer vision applications. PD results from factors such as camera positioning, depth, intrinsic parameters like focal length and lens distortion, and extrinsic parameters like rotation and translation. These factors collectively impact the projection of 3D scenes onto 2D planes, affecting semantic interpretation and local geometry. Accurately estimating these parameters for PD correction is difficult, posing a major obstacle to creating robust computer vision (CV) methods.

We recommend citing following publication for using ImageNet-PD:

Chhipa, P.C., Chippa, M.S., De, K., Saini, R., Liwicki, M., Shah, M.: Möbius transform for mitigating perspective distortions in representation learning. European Conference on Computer Vision (2024).

Two minutes summary on MPD method and links to access source code repository and MPD pretrained models are available at https://prakashchhipa.github.io/projects/mpd/.

Chhipa, P. C., Chippa, M. S., De, K., Saini, R., Liwicki, M., & Shah, M. (2024). M" obius Transform for Mitigating Perspective Distortions in Representation Learning. arXiv preprint arXiv:2405.02296.