Large vision-language representation learning models like CLIP have demonstrated impressive performance for zero-shot transfer to downstream tasks while largely benefiting from inter-modal (image-text) alignment via contrastive objectives. This downstream performance can further be enhanced by full-scale fine-tuning which is often compute intensive, requires large labelled data, and can reduce out-of-distribution (OOD) robustness. Furthermore, sole reliance on inter-modal alignment might overlook the rich information embedded within each individual modality. In this work, we introduce a sample-efficient domain adaptation strategy for CLIP, termed Domain Aligned CLIP (DAC), which improves both intra-modal (image-image) and inter-modal alignment on target distributions without fine-tuning the main model. For intra-modal alignment, we introduce a lightweight adapter that is specifically trained with an intra-modal contrastive objective. To improve intermodal alignment, we introduce a simple framework to modulate the precomputed class text embeddings. The proposed few-shot fine-tuning framework is computationally efficient, robust to distribution shifts, and does not alter CLIP’s parameters. We study the effectiveness of DAC by benchmarking on 11 widely used image classification tasks with consistent improvements in 16-shot classification upon strong baselines by about 2.3% and demonstrate competitive performance on 4 OOD robustness benchmarks.

This work presents a novel approach for semi-supervised semantic segmentation. The key element of this approach is our contrastive learning module that enforces the segmentation network to yield similar pixel-level feature representations for same-class samples across the whole dataset. To achieve this, we maintain a memory bank which is continuously updated with relevant and high-quality feature vectors from labeled data. In an end-to-end training, the features from both labeled and unlabeled data are optimized to be similar to same-class samples from the memory bank. Our approach not only outperforms the current state-of-the-art for semi-supervised semantic segmentation but also for semi-supervised domain adaptation on well-known public benchmarks, with larger improvements on the most challenging scenarios, i.e., less available labeled data. Code is available at https://github.com/Shathe/SemiSeg-Contrastive

LiDAR semantic segmentation provides 3D semantic information about the environment, an essential cue for intelligent systems, such as autonomous vehicles, during their decision making processes. Unfortunately, the annotation process for this task is very expensive. To overcome this, it is key to find models that generalize well or adapt to additional domains where labeled data is limited. This work addresses the problem of unsupervised domain adaptation for LiDAR semantic segmentation models. We propose simple but effective strategies to reduce the domain shift by aligning the data distribution on the input space. Besides, we present a learning-based module to align the distribution of the semantic classes of the target domain to the source domain. Our approach achieves new state-of-the-art results on three different public datasets, which showcase adaptation to three different domains.

Efficient models for semantic segmentation, in terms of memory, speed, and computation, could boost many robotic applications with strong computational and temporal restrictions. This article presents a detailed analysis of different techniques for efficient semantic segmentation. Following this analysis, we have developed a novel architecture, MiniNet-v2, an enhanced version of MiniNet. MiniNet-v2 is built considering the best option depending on CPU or GPU availability. It reaches comparable accuracy to the state-of-the-art models but uses less memory and computational resources. We validate and analyze the details of our architecture through a comprehensive set of experiments on public benchmarks (Cityscapes, Camvid, and COCO-Text datasets), showing its benefits over relevant prior work. Our experiments include a sample application where these models can boost existing robotic applications. Alonso, Íñigo; Riazuelo, Luis; Murillo, Ana C.

Selecting relevant visual information from a video is a challenging task on its own and even more in robotics, due to strong computational restrictions. This work proposes a novel keyframe selection strategy based on image quality and semantic information, which boosts strategies currently used in Visual-SLAM (V-SLAM). Commonly used V-SLAM methods select keyframes based only on relative displacements and amount of tracked feature points. Our strategy to select more carefully these keyframes allows the robotic systems to make better use of them. With minimal computational cost, we show that our selection includes more relevant keyframes, which are useful for additional posterior recognition tasks, without penalizing the existing ones, mainly place recognition. A key ingredient is our novel CNN architecture to run a quick semantic image analysis at the onboard CPU of the robot. It provides sufficient accuracy significantly faster than related works. We demonstrate our hypothesis with several public datasets with challenging robotic data.

Semantic segmentation is a challenging problem that can benefit numerous robotics applications, since it provides information about the content at every image pixel. Solutions to this problem have recently witnessed a boost on performance and results thanks to deep learning approaches. Unfortunately, common deep learning models for semantic segmentation present several challenges which hinder real life applicability in many domains. A significant challenge is the need of pixel level labeling on large amounts of training images to be able to train those models, which implies a very high cost. This work proposes and validates a simple but effective approach to train dense semantic segmentation models from sparsely labeled data. Labeling only a few pixels per image reduces the human interaction required. We find many available datasets, e.g., environment monitoring data, that provide this kind of sparse labeling. Our approach is based on augmenting the sparse annotation to a dense one with the proposed adaptive superpixel segmentation propagation. We show that this label augmentation enables effective learning of state-of-the-art segmentation models, getting similar results to those models trained with dense ground-truth. We demonstrate the applicability of the presented approach to different image modalities in real domains (underwater, aerial and urban scenarios) with publicly available datasets.