|
I am a senior engineer at Qualcomm AI Research, working on deep learning for efficient video processing with Amirhossein Habibian. I got my Ph.D. at AimageLab, at the University of Modena and Reggio Emilia, in Italy. I worked under the supervision of Prof. Rita Cucchiara. Email / CV / Google Scholar / Github |
|
|
I am currently working on efficient video processing, investigating techniques such as conditional computation and distillation to speed up operations over sequences of frames. In my previous Ph.D. life, I have worked on computer vision for automotive applications, out-of-distribution detection and continual learning. A full publication list can be found on my scholar page. |
 |
Kumara Kahatapitiya, Adil Karjauv, Davide Abati, Fatih Porikli, Yuki Asano, Amirhossein Habibian ECCV, 2024 arXiv / project page / bibtex Reducing computational cost of diffusion-based video editing methods by 10x, by squeezing operations on unedited images at the minimum. |
 |
Davide Abati, Haitam Ben Yahia, Markus Nagel, Amirhossein Habibian ICCV, 2023 arXiv / bibtex When processing a video, residuals in frame representations can be processed at a very low integer precision, with very low quantization error. |
 |
Yura Perugachi Diaz, Guillaume Sautière, Davide Abati, Yang Yang, Amirhossein Habibian, Taco Cohen BMVC, 2022 (Oral presentation) arXiv / bibtex Enabling neural based video codecs to improve distortion within regions of interest, while spending less bits on background areas. |
 |
Amirhossein Habibian, Haitam Ben Yahia, Davide Abati, Efstratios Gavves, Fatih Porikli ECCV, 2022 arXiv / bibtex / code Instead of distilling neural networks activations, we teach a student network to regress their temporal differences, allowing an improved cost-performance tradeoff. |
 |
Davide Abati, Amir Ghodrati, Amirhossein Habibian BMVC, 2021 bibtex An efficient frame alignment operator for Video Super Resolution based on local self attention. Alignment is carried out only at specific locations as specified by a learnable gating function. |
 |
Amirhossein Habibian, Davide Abati, Taco Cohen, Babak Ehteshami Bejnordi CVPR, 2021 arXiv / bibtex / code By selectively applying convolutions only on locations that carry meaningful information, the computational cost of neural networks on video can be reduced by 4~5 times. |
 |
Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, Simone Calderara NeurIPS, 2020 arXiv / bibtex / code We tackle general continual learning, where an agent is required to learn multiple tasks in a sequence under minimal assumptions about their nature. We run an extensive comparison of many existing methods and introduce a simple model based on knowledge distillation outperforming all of them. |
 |
Davide Abati, Jakub Tomczak, Tijmen Blankevoort, Simone Calderara, Rita Cucchiara, Babak Ehteshami Bejnordi CVPR, 2020 (Oral presentation) arXiv / bibtex / talk / lecture A continual learning model based on the framework of conditional computation. |
 |
Davide Abati, Angelo Porrello, Simone Calderara, Rita Cucchiara CVPR, 2019 arXiv / code / poster / bibtex Applying autoregression in an autoencoder's latent space increases its out-of-distribution detection capabilities. |
 |
Andrea Palazzi, Davide Abati, Francesco Solera, Simone Calderara, Rita Cucchiara IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018 arXiv / dataset / code / video / bibtex We introduce a dataset of human fixations while driving, and a model to predict them given an urban scene. |
|
Journals:
Conferences:
|
|
|
|