Orgulloso de mi pais en estos momentos tan difíciles 🇪🇸 Proud of my country in these extremely hard moments. All my heart is with mi city and all the victims from this tragedy :( 💔.

If you are at #ECCV and want to know about calibrating large language-vision model adaptors, come now to our poster (nr 79)

RT @DLMI2024: All about Foundation Models today at the last day of summer school #DLMI with Prof @josedolz_ets! . #ChatGPT #FoundationModel….

RT @DLMI2024: We begin our day with a talk on Weakly supervised #deeplearning , constrained losses and semantic segmentation! . #DLMI2024 h….

RT @DLMI2024: Poster sessions! #DLMI2024

RT @DLMI2024: We have already begun the summer school! .#DLMI2024

so true. lol.

I am hiring two post-docs to work at the intersection of medical imaging and machine learning (modelling the uncertainty of large language-vision models). If you are interested, drop me an email for more information jose.dolz@etsmtl.ca.

If you are interested in doing a PhD in deep learning and computer vision at ETS Montreal, drop me an email with your CV and research interests (more info in the attached image)

RT @mertrory: Journal -> Conference!.

5/5. Link to the paper: Github: Congrats to all co-authors!! @bing_bingyuan @adrian_galdran @IsmailBenAyed1.

4/5. Dice has an intrinsic bias towards specific extremely imbalanced solutions, whereas CE implicitly encourages the ground-truth region proportions. This explains the wide experimental evidence in medical-imaging, where Dice loss brings improvements for imbalanced segmentation.

3/5.And the second one, a region-size penalty term imposing different biases on the size (or proportion) of the predicted regions. Our information-theoretic analysis uncovers hidden region-size biases.

2/5. In this work, we provide a theoretical analysis, which shows that CE and Dice share a deep connection. They both decompose into two components. The first one, a similar ground-truth matching term, which pushes the predicted foreground regions towards the ground-truth;.

1/5.Do you wonder which is the best loss function to use in your medical segmentation model? It is widely argued within the medical-imaging community that Dice and CE losses are complementary, which has motivated the use of compound CE-Dice losses (the de-facto solution nowadays).

🚨One of our latest papers, where we propose to use Denoising Auto-Encoders to model the uncertainty of the predictions in semi-supervised segmentation has been accepted in MedIA. Congrats @sukeshadiga !! 🎉🎉💪💪.Arxiv: Github:

Kudos to all the first authors and co-authors on these papers!! @93Balamuralim @jul_nicol @IsmailBenAyed1 @imtiaz_masud.

2) MoP-CLIP: A Mixture of Prompt-Tuned CLIP Models for Domain Incremental Learning. Paper:

It seems some of my students will go to Hawaii this winter to present their works at @wacv_official .1) Prompting classes: Exploring the Power of Prompt Class Learning in Weakly Supervised Semantic Segmentation. Paper: Github:

Email to: jose.dolz@etsmtl.ca.