Competition (Advanced AI, Fall 2025)

Topic : Image Deblurring

Introduction : Image deblurring is a very interesting technique in the field of computer vision, which restores the sharp information from the blurry image. As students expected, this problem is quite under constrained, thus previous approaches in literature attempted to resolve this problem via the large amount of training pairs (i.e., blurry-sharp pairs), which are acquired based on the special hardware setting. In this project, we will apply the convolutional neural network to resolve the problem of image deblurring in an automatic manner. Through this project, students have a good opportunity to implement the convolutional neural network for the real-world application by yourself. Some examples of image deblurring are given as follows:

Dataset : For this competition, students need to use "GoPro Dataset [1]", which is provided by SNU. Note that you can define your training set and validation set by using all the samples of the GoPro dataset (3,214 blurred images)

• [1] S. Nah, T. H. Kim, and K. M. Lee, "Deep multi-scale convolutional neural network for dynamic scene deblurring," in Proc. IEEE Int. Conf. Comput. Vis. Pattern Recognit., pp. 3883-3891, Jun. 2017. (Dataset link : https://seungjunnah.github.io/Datasets/gopro)

• You need to re-categorize it into two sets as follows:

  • Training set : you need to define it by yourself (based on samples in the GoPro dataset)

  • Validation set : you need to define it by yourself (based on samples in the GoPro dataset)

  • Test set : Test samples will be given on 1. Dec. (21:00) (see the Schedule Section for more details) 

Performance Evaluation : Students train your networks based on the training set. To tune your networks (weights as well as network architectures), you can use the validation samples. After that, we provide our test samples for the performance evaluation later. As the evaluation criterion, we use the PSNR metric (it can be regarded as mean square error in the log domain).

Framework : Google Colab (if you already have GPU on your local PC, then it is OK to use it) 

Policy : Students must design their network architectures by themselves based on contents they have learned in class (or their own idea) for this competition. 🚨 Any other use of open sources for image deblurring is prohibited (if you use them, credits are not given).

Schedule :  Test samples will be given on 1. Dec. (21:00). Then, students should submit their predicted results (i.e., result images) within 30 minutes. After that, the performance by the PSNR metric will be reported ASAP.

Results of the performance evaluation :  I will report the summary of this competition in class 3. Dec.

Final submission (최종 제출물) :  1) Report (including your architecture as a figure), 2) Source code, 3) Result images

Due : 1. Dec. (21:30) submit yours by using e-campus "assignment" tap (과제 탭)

🎉 Students who submit meaningful results will be credited (up to) 10 points in Final Exam (or Mid-term Exam if you get > 90 in Final Exam).