LAMP

The Learning and Machine Perception (LAMP) team at the Computer Vision Center conducts fundamental research and technology transfer in the field of machine learning for semantic understanding of visual data. The group works with a wide variety of visual data sources: from multispectral, medical imagery and consumer camera images, to live webcam streams and video data. The returning objective is the design of efficient and accurate algorithms for the automatic extraction of semantic information from visual media.

Three papers were accepted:
1. Elastic Feature Consolidation for Cold Start Exemplar-free Incremental Learning (pdf).
2. Get What You Want, Not What You Don’t: Image Content Suppression for Text-to-Image Diffusion Models (pdf).
3. Divide and not forget: Ensemble of selectively trained experts in Continual Learning (pdf).

Two papers were accepted:
1. FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning (pdf).
2. Dynamic Prompt Learning: Addressing Cross-Attention Leakage for Text-Based Image Editing (pdf).

And one workshop paper:
1.IterInv: Iterative Inversion for Pixel-Level T2I Models (pdf)

Two papers were accepted:

1. Augmented Box Replay: Overcoming Foreground Shift for Incremental Object Detection (pdf).
2. ICICLE: Interpretable Class Incremental Continual Learning (pdf).

Our paper on Planckian Jitter for better color image representation has been accepted for ICLR. Great work Simone and Alex!

1. Planckian Jitter: countering the color-crippling effects of color jitter on self-supervised training (pdf).

Shiqi-s paper Attracting and Dispersing: A Simple Approach for Source-free Domain Adaptation got accepted for NeurIPS 2022.

The survey paper Class-incremental learning: survey and performance evaluation on image classification is accepted at PAMI.

Also check the code framework FACIL that allows to reproduce the results from the survey.

The paper on zero-shot has also been accepted a PAMI: Generative Multi-Label Zero-Shot Learning

And in IJCV: MineGAN++: Mining Generative Models for Efficient Knowledge Transfer to Limited Data Domains

Alex won the Best Paper Award at the Continual Learning Workshop at CVPR 2022 for his paper Continually Learning Self-Supervised Representations with Projected Functional Regularization. Francesco and Saurav received the runner-up award for:
Towards Exemplar-Free Continual Learning in Vision Transformers: an Account of Attention, Functional and Weight Regularization.

Joost van de Weijer gave an invited talk at Continual Learning Workshop.

Shiqi has one paper at the NeurIPS 2021 Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation.

Riccardo’s paper >RATT: Recurrent Attention to Transient Tasks for Continual Image Captioning on continual learning of captioning systems, and Yaxing’s paper on transfer learning for image-to-image systems:DeepI2I: Enabling Deep Hierarchical Image-to-Image Translation by Transferring from GANs have been accepted !

Four CVPR papers have been accepted:

• MineGAN: effective knowledge transfer from GANs to target domains with few images,
• Orderless Recurrent Models for Multi-label Classification,
• Semantic Drift Compensation for Class-Incremental Learning,
• Semi-supervised Learning for Few-shot Image-to-Image Translation.

and one workshop paper:

• Generative Feature Replay For Class-Incremental Learning,