DTU.dk

Computer Vision

Computer vision is one the core research fields of the Visual Computing section at DTU Compute.

Our highlights include

Visual data restoration

Statistical image modeling

Compression

Visual recognition with minimal supervision

Multimodal learning

Camera calibration

3D scanning

Visual data restoration

Many problems in computer vision involve conditional image generation, acquisition and display.. Deep learning has greatly advanced the quality of the generated images in recent years; however challenges still remain. These problems often come with paired training data, making the image generation supervised. We develop innovative image restoration techniques using deep neural networks. These techniques can solve many problems such as image denoising, demosaicing, super-resolution, and inpainting. We also work on video frame interpolation which is able to increase the temporal resolution of a video.

For further information see:

Contact

Dimitrios Papadopoulos

Associate Professor

DTU Compute

dimp@dtu.dk

Siavash Arjomand Bigdeli

Associate Professor

DTU Compute

sarbi@dtu.dk

Morten Rieger Hannemose

Assistant Professor

DTU Compute

mohan@dtu.dk

Statistical image modeling

We develop deep learning models for data density estimation with applications in out-of-distribution detection and compression. We also build generative models for various visual data such as face images and 3D textures.

For further information see:

Bigdeli, Siavash A., Geng Lin, Tiziano Portenier, L. Andrea Dunbar, and Matthias Zwicker. "Learning generative models using denoising density estimators." arXiv preprint arXiv:2001.02728 (2020).

Portenier, Tiziano, Qiyang Hu, Attila Szabo, Siavash Arjomand, Paolo Favaro, and Matthias Zwicker. "FaceShop: Deep Sketch-based Image Editing." ACM transactions on graphics 37, no. 4 (2018): 1-13.

Portenier, Tiziano, Siavash Arjomand Bigdeli, and Orcun Goksel. "Gramgan: Deep 3d texture synthesis from 2d exemplars." Advances in Neural Information Processing Systems 33 (2020): 6994-7004.

Compression

We are investigating deep learning-based AI techniques to address data compression challenges. We are working with high dimensional datasets as well as 2D natural images. We also look at innovative neural network approaches to reduce their size with respect to their performance.

For further information see:

Bigdeli, Siavash A., Geng Lin, Tiziano Portenier, L. Andrea Dunbar, and Matthias Zwicker. "Learning generative models using denoising density estimators." arXiv preprint arXiv:2001.02728 (2020).

Narduzzi, Simon, Siavash A. Bigdeli, Shih-Chii Liu, and L. Andrea Dunbar. "Optimizing the consumption of spiking neural networks with activity regularization." In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 61-65. IEEE, 2022.

Visual recognition with minimal supervision

A fundamental component of most visual recognition systems is the learning procedure. Like humans, the machine needs to be trained beforehand with a lot of training examples to be able to learn visual models and make predictions in unseen data. Obtaining such data typically requires human annotation which is tedious, very expensive and time consuming. To alleviate this problem, we focus on developing efficient techniques for learning visual models with minimal supervision for the tasks of object detection and image segmentation. We have proposed weakly supervised models with efficient human interaction, human-in-the-loop schemes and label propagation techniques for efficient image annotation.

For further information see:

Multimodal learning

Humans and even animals use inputs from different senses for recognizing or understanding something and take decisions accordingly. Similarly, machines in real-world applications often use multiple sensors that produce multi-modal data. We have developed multimodal systems that combine vision and language for several domains such as artificial intelligence for social good or food recognition.

For further information see:

Camera calibration

Having an accurately calibrated camera setup (intrinsics and extrinsics) is essential for many computer vision problems such as 3D scanning. We develop practical methods for highly accurate camera calibration

For further information see:

Hannemose, Morten, Jakob Wilm, and Jeppe Revall Frisvad. "Superaccurate camera calibration via inverse rendering." Modeling Aspects in Optical Metrology VII. Vol. 11057. SPIE, 2019..

3D scanning

Our lab has extensive experience in structured light 3D scanning. Recent improvements in projector technology, increased processing power, and new method developments with central contributions from our research group, it is now possible to perform faster and highly accurate structured light scans. This offers new opportunities for studying dynamic scenes, quality control, human-computer interaction and more.

Courses

Computer vision is a central topic in the following DTU courses

External references

We also collaborate and are a part of the DTU 3D Imaging Center - 3DIM, which is our X-ray µCT laboratory.

Updated by Mechid Paiman on 30 April 2023