About the Turbulence Simulator

Due to the difficulties in collecting distorted/clean image pairs, researchers in the field of atmospheric turbulence mitigation are mostly working with synthetic data. The P2S simulator is the latest atmospheric turbulence simulator developed by Purdue University. It is, to our knowledge, the first physical-based simulator that can achieve close to real-time speed for simulating the atmospheric turbulence effect on images.

Paper and simulator code: paper, code

Evaluation

Subtrack1: Text Recognition through Atmospheric Turbulence.

In the dry run stage, there are 50 sequences of text images for validation. Each sequence will contain 100 turbulence distorted frames. The task for the participants is to reconstruct a high quality image from these 100 distorted frames (you can use all or part of these frames). The ground truth labels are provided. The evaluation metrics of this phase is the average accuracy of three existing scene text recognition algorithms on the reconstruction result of the text patterns.

In the final testing stage, there will be 500 sequences of text images in the same format as the dry run phase. Each sequence will contain 100 turbulence distorted frames. No ground truth labels will be available to participants. The ranking in this subtrack will be based on the average accuracy of three existing scene text recognition algorithms on the reconstruction result of the text patterns.

Get dry run data and starting kit.

Get final testing data .

Subtrack2: Coded Target Restoration through Atmospheric Turbulence

In the dry run stage, there are 32 sequences of code images for validation. Each sequence will contain 100 turbulence distorted frames. The task for the participants is also to reconstruct a single high quality image from these 100 distorted frames (you can use all or part of these frames). No ground truth labels are provided and we expect participants to upload their restored results directly. The evaluation metrics of this phase is the average bit score from the reconstruction result of the code patterns.

In the final testing stage, there will be 192 sequences of code images in the same format as the dry run phase. Each sequence will contain 100 turbulence distorted frames. The ranking in this subtrack will be based on the average bit score estimated from the reconstruction result.

Get dry run data.

Get final testing data.

If you have any questions about this challenge track please feel free to email cvpr2023.ug2challenge@gmail.com