Existing emotion prediction benchmarks contain coarse emotion labels which do not consider the diversity of emotions that an image and text can elicit in humans due to various reasons. Learning diverse reactions to multimodal content is important as intelligent machines take a central role in generating and delivering content to society. To address this gap, we propose Socratis, a societal reactions benchmark, where each image-caption (IC) pair is annotated with multiple emotions and the reasons for feeling them. Socratis contains 18K free-form reactions for 980 emotions on 2075 image-caption pairs from 5 widely-read news and image-caption (IC) datasets. We benchmark the capability of state-of-the-art multimodal large language models to generate the reasons for feeling an emotion given an IC pair. Based on a preliminary human study, we observe that humans prefer human-written reasons over 2 times more often than machine-generated ones. This shows our task is harder than standard generation tasks because it starkly contrasts recent findings where humans cannot tell apart machine vs human-written news articles, for instance. We further see that current captioning metrics based on large vision-language models also fail to correlate with human preferences. We hope that these findings and our benchmark will inspire further research on training emotionally aware models.
Visual recognition models are prone to learning spurious correlations induced by an imbalanced training set where certain groups (\eg Females) are under-represented in certain classes (\eg Programmers). Generative models offer a promising direction in mitigating this bias by generating synthetic data for the minority samples and thus balancing the training set. However, prior work that uses these approaches overlooks that visual recognition models could often learn to differentiate between real and synthetic images and thus fail to unlearn the bias in the original dataset. In our work, we propose a novel two-stage pipeline to mitigate this issue where 1) we pre-train a model on a balanced synthetic dataset and then 2) fine-tune on the real data. Using this pipeline, we avoid training on both real and synthetic data, thus avoiding the bias between real and synthetic data. Moreover, we learn robust features against the bias in the first step that mitigate the bias in the second step. Moreover, our pipeline naturally integrates with bias mitigation methods; they can be simply applied to the fine-tuning step. As our experiments prove, our pipeline can further improve the performance of bias mitigation methods obtaining state-of-the-art performance on three large-scale datasets.
Long-term activity forecasting is an especially challenging research problem because it requires understanding the temporal relationships between observed actions, as well as the variability and complexity of human activities. Despite relying on strong supervision via expensive human annotations, state-of-the-art forecasting approaches often generalize poorly to unseen data. To alleviate this issue, we propose Multiscale Video Pretraining (MVP), a novel self-supervised pretraining approach that learns robust representations for forecasting by learning to predict contextualized representations of future video clips over multiple timescales. MVP is based on our observation that actions in videos have a multiscale nature, where atomic actions typically occur at a short timescale and more complex actions may span longer timescales. We compare MVP to state-of-the-art self-supervised video learning approaches on downstream long-term forecasting tasks including long-term action anticipation and video summary prediction. Our comprehensive experiments across the Ego4D and Epic-Kitchens-55/100 datasets demonstrate that MVP out-performs state-of-the-art methods by significant margins. Notably, MVP obtains a relative performance gain of over 20% accuracy in video summary forecasting over existing methods.
Learning with noisy labels (LNL) is challenging as the model tends to memorize noisy labels, which can lead to overfitting. Many LNL methods detect clean samples by maximizing the similarity between samples in each category, which does not make any assumptions about likely noise sources. However, we often have some knowledge about the potential source(s) of noisy labels. For example, an image mislabeled as a cheetah is more likely a leopard than a hippopotamus due to their visual similarity. Thus, we introduce a new task called Learning with Noisy Labels and noise source distribution Knowledge (LNL+K), which assumes we have some knowledge about likely source(s) of label noise that we can take advantage of. By making this presumption, methods are better equipped to distinguish hard negatives between categories from label noise. In addition, this enables us to explore datasets where the noise may represent the majority of samples, a setting that breaks a critical premise of most methods developed for the LNL task. We explore several baseline LNL+K approaches that integrate noise source knowledge into state-of-the-art LNL methods across three diverse datasets and three types of noise, where we report a 5-15% boost in performance compared with the unadapted methods. Critically, we find that LNL methods do not generalize well in every setting, highlighting the importance of directly exploring our LNL+K task.
Diffusion models have demonstrated impressive performance in text-guided image generation. To leverage the knowledge of text-guided image generation models in image editing, current approaches either fine-tune the pretrained models using the input image (e.g., Imagic) or incorporate structure information as additional constraints into the pretrained models (e.g., ControlNet). However, fine-tuning large-scale diffusion models on a single image can lead to severe overfitting issues and lengthy inference time. The information leakage from pretrained models makes it challenging to preserve the text-irrelevant content of the input image while generating new features guided by language descriptions. On the other hand, methods that incorporate structural guidance (e.g., edge maps, semantic maps, keypoints) as additional constraints face limitations in preserving other attributes of the original image, such as colors or textures. A straightforward way to incorporate the original image is to directly use it as an additional control. However, since image editing methods are typically trained on the image reconstruction task, the incorporation can lead to the identical mapping issue, where the model learns to output an image identical to the input, resulting in limited editing capabilities. To address these challenges, we propose a text-to-image editing model with Image Information Removal module (IIR) to selectively erase color-related and texture-related information from the original image, allowing us to better preserve the text-irrelevant content and avoid the identical mapping issue. We evaluate our model on three benchmark datasets: CUB, Outdoor Scenes, and COCO. Our approach achieves the best editability-fidelity trade-off, and our edited images are approximately 35% more preferred by annotators than the prior-arts on COCO.
Webpages have been a rich resource for language and vision-language tasks. Yet only pieces of webpages are kept: image-caption pairs, long text articles, or raw HTML, never all in one place. Webpage tasks have resultingly received little attention and structured image-text data underused. To study multimodal webpage understanding, we introduce the Wikipedia Webpage 2M (WikiWeb2M) suite; the first to retain the full set of images, text, and structure data available in a page. WikiWeb2M can be used for tasks like page description generation, section summarization, and contextual image captioning.
Compositional reasoning is a hallmark of human visual intelligence; yet despite the size of large vision-language models, they struggle to represent simple compositions by combining objects with their attributes. To measure this lack of compositional capability, we design Cola, a text-to-image retrieval benchmark to Compose Objects Localized with Attributes. Using Cola as a testbed, we explore modeling designs to adapt pre-trained vision-language models to reason compositionally about multiple attributes attached to multiple objects. We explore 6 finetuning strategies on 2 seminal vision-language models, using 3 finetuning datasets and 2 test benchmarks (Cola and CREPE). Surprisingly, our optimal finetuning strategy improves a 151M parameter CLIP, which disjointly encodes image and language during pretraining, to perform as well as a 241M parameter FLAVA, which uses a multi-modal transformer encoder during pretraining to attend over both vision and language modalities. This optimal finetuning strategy is a lightweight multi-modal adapter that jointly attends over both image and language features generated by the pretrained model. We show this works better than common strategies such as prompt/fine-tuning, or tuning a comparable number of unimodal layers.
Webpages have been a rich, scalable resource for vision-language and language only tasks. Yet only pieces of webpages are kept: image-caption pairs, long text articles, or raw HTML, never all in one place. Webpage tasks have resultingly received little attention and structured image-text data left underused. To study multimodal webpage understanding, we introduce the Wikipedia Webpage suite (WikiWeb2M) of 2M pages. We verify its utility on three generative tasks: page description generation, section summarization, and contextual image captioning. We design a novel attention mechanism Prefix Global, which selects the most relevant image and text content as global tokens to attend to the rest of the webpage for context. By using page structure to separate such tokens, it performs better than full attention with lower computational complexity. Experiments show that the new annotations from WikiWeb2M improve task performance compared to data from prior work. We also include ablations on sequence length, input features, and model size.
Multi-source Domain Generalization (DG) measures a classifier's ability to generalize to new distributions of data it was not trained on, given several training domains. While several multi-source DG methods have been proposed, they incur additional complexity during training by using domain labels. Recent work has shown that a well-tuned Empirical Risk Minimization (ERM) training procedure, that is simply minimizing the empirical risk on the source domains, can outperform most existing DG methods. We identify several key candidate techniques to further improve ERM performance, such as better utilization of training data, model parameter selection, and weight-space regularization. We call the resulting method ERM++, and show it significantly improves the performance of DG on five multi-source datasets by over 5% compared to standard ERM, and beats state-of-the-art despite being less computationally expensive. Additionally, we demonstrate the efficacy of ERM++ on the WILDS-FMOW dataset, a challenging DG benchmark. We hope that ERM++ becomes a strong baseline for future DG research. Code is released at https://github.com/piotr-teterwak/erm_plusplus.