Learning Aligned Cross-Modal Representations from Weakly Aligned Data

By Lluís Castrejón, Yusuf Aytar, Carl Vondrick, Hamed Pirsiavash, Antonio Torralba. 2016

183 citations recorded by the source corpus.

People can recognize scenes across many different modalities beyond natural images. In this paper, we investigate how to learn cross-modal scene representations that transfer across modalities. To study this problem, we introduce a new cross-modal scene dataset. While convolutional neural networks can categorize cross-modal scenes well, they also learn an intermediate representation not aligned across modalities, which is undesirable for crossmodal transfer applications. We present methods to regularize cross-modal convolutional neural networks so that they have a shared representation that is agnostic of the modality. Our experiments suggest that our scene representation can help transfer representations across modalities for retrieval. Moreover, our visualizations suggest that units emerge in the shared representation that tend to activate on consistent concepts independently of the modality.

Subfields: Computer Vision and Pattern Recognition, Multimodal Learning, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Multimodal Learning

In Representation & Self-Supervised Learning

Citations in the corpus

The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation (2016, background): “More recently, it has been shown that deep networks can also boost their performance when pretrained on data other than natural images, such as video games [26, 28] or clipart [3], and this an interesting direction to explore.”
SoundNet: Learning Sound Representations from Unlabeled Video (2016, background): “Cross-Modal Learning and Unlabeled Video: Our approach is broadly inspired by efforts to model cross-modal relations [24, 14, 7, 26] and works that leverage large amounts of unlabeled video [25, 41, 8, 40, 39].”
GuessWhat?! Visual Object Discovery through Multi-modal Dialogue (2016, background): “dence interval before proceeding to the ﬁnal guess. Differently, GuessWhat?! could be a test bed for one-shot learning [14] of guessing new object categories, transfer learning on line-drawing images [10] or using questions from another language. Thus, GuessWhat?! dataset offers a opportunity to dev”
Domain Adaptation for Visual Applications: A Comprehensive Survey (2017, methodology): “ation, pose changes, but the domain mismatch might be more severe when, for example, the source and target domains contain images of different types, such as photos, NIR images, paintings or sketches [1, 2, 3, 4]. Service provider companies are especially concerned since, for the same service (task)”
Through-Wall Human Pose Estimation Using Radio Signals (2018, background): “In terms of modeling, our work is related to cross-modal and multi-modal learning that explores matching different modalities or delivering complementary information across modalities [8, 11, 36, 34].”
Dual-path Convolutional Image-Text Embeddings with Instance Loss (2017, background): “deploy the multiple labels to learn the shared semantic space [1].”
Learning Visually-Grounded Semantics from Contrastive Adversarial Samples (2018, background): “…for shape inference, bilingual word embeddings for machine translation (Zou et al., 2013), human pose-image embeddings for pose inference (Li, 2011), image-text embeddings for visual description (Reed et al., 2016), and global representation learning from multiple domains (Castrejon et al., 2016).”
MultiMAE: Multi-modal Multi-task Masked Autoencoders (2022, methodology): “It can either involve training separate encoders or one unified architecture (e.g., a Transformer [85]) to operate on modalities such as images and text [4, 14, 18, 39, 42–45, 56, 57, 76, 79, 94], video and audio [5,41,60,62], video, text and audio [3], and depth, images and video [32].”
Multimodal MR Synthesis via Modality-Invariant Latent Representation (2017, background): “Thus, our approach is broadly similar to the statistical regularisation approach in [32], in which cross-modal scene representations are learnt.”
Context Autoencoder for Self-supervised Representation Learning (2022, unknown): “We conduct fine-tuning experiments on three datasets: Food-101 [6], Clipart [12], and Sketch [12].”
Deep Partial Multi-View Learning (2020, methodology): “The cross-modal convolutional neural networks are regularized to obtain a shared representation which is agnostic to the modality for cross-modal scene images [31].”
PRNet: Self-Supervised Learning for Partial-to-Partial Registration (2019, methodology): “Inspired by this observation, self-supervised learning usually involves predicting part of an input from another part [47, 48], solving one task using features learned from another task [45] and/or enforcing consistency from different views/modalities [49, 50].”
Who is Mistaken? (2016, unknown): “dditionally, we extend this previous work to multi-frame scenes depicting visual stories. Transfer: After we learn to recognize mistaken characters in abstract scenes, one could use domain adaptation [12,6] to apply our approach to natural images. However, this is orthogonal to the goal of this pape”
Label Efficient Learning of Transferable Representations acrosss Domains and Tasks (2017, result): “Weight sharing of CNN parameters [58, 22, 21, 3] and minimizing the distribution discrepancy of network activations [51, 65, 30] have also shown convincing results.”
Visually Grounded Neural Syntax Acquisition (2019, unknown): “Its applicability could be extended by learning shared representations across multiple modalities (Castrejon et al., 2016) or integrating with pure text-domain models (such as PRPN, Shen et al., 2018a).”
Cross-Domain Weakly-Supervised Object Detection Through Progressive Domain Adaptation (2018, background): “, CMPlaces [4]) and two image search engines (i.”
Automatic Understanding of Image and Video Advertisements (2017, background): “There is work in learning domaininvariant representations [22, 75, 7, 8, 18, 20, 31, 43, 11], but a challenge in the case of ads is that data from each “domain” (the particular way e.”
A Comprehensive Survey on Cross-modal Retrieval (2016, background): “Castrejon et al. [97] present two approaches to regularize cross-modal convolutional networks so that the intermediate representations are aligned across modalities.”