Abstract:
Building robust, generalizable models that work in the wild across many applications requires learning that does not rely on manual labels. Human annotations are costly and time-consuming. Labels inherit human biases and conventions, reflect taxonomic choices and annotation artifacts, and cannot be obtained at scale or keep pace with an always-changing, long-tailed world. Meanwhile, vast unlabeled data across many modalities (e.g., video, audio, text) are freely available online. Humans naturally learn without explicit teaching by discovering regularities in unimodal and multimodal inputs, integrating information across modalities, and using language to reason about them. This thesis advances a label-free path for image and video understanding, building on ideas of multimodal learning, reasoning in language space, and unsupervised learning.
This thesis contributes to different main stages of building vision models without labels: from creating high-quality datasets for pretraining, to training strong general-purpose models, adapting them to specific tasks and domains, exploring models' dense properties, and finally improving evaluation benchmarks -- all without any human annotations. We develop three complementary ideas: (1) leverage multimodal pretraining to go beyond simple modality-alignment by learning fused multimodal representations, enabling dense spatio-temporal grounding capabilities, and using multimodal models as pseudo-supervisors for task/domain adaptation; (2) use language as the powerful operating space by translating visual signals into text descriptions and exploiting the strong reasoning abilities of large language models; and (3) design losses, architectures, and pipelines for unsupervised learning across different stages, such as pretraining, domain/task adaptation, and spatio-temporal grounding. Concretely, we introduce an ordering-based contrastive objective for self-supervised training that improves the local neighborhood structure of the learned embedding space compared to a commonly used contrastive loss. We then construct truly multimodal video–audio–text representations that can flexibly fuse any subset of modalities into a single joint embedding. Then, we present a pipeline that adapts image–language models to activity-centric video understanding, namely action recognition, using only unlabeled videos together with action dictionaries, eliminating the need for paired supervision. We also develop a pipeline that adapts vision–language models to specific domains (e.g., cooking) using only domain text data, without even target-domain vision data or any paired annotations. To increase the quality of of web-scale video–audio–text data for multimodal pretraining without human curation, we propose a method to convert noisy subtitles of web instructional videos into dense, high-quality, human-like descriptions by using LLM to reason about the videos entirely in language space. Using the same language-space reasoning idea, we propose a method to diagnose and mitigate representation biases, such as object or single-frame bias, in popular video evaluation benchmarks, enabling a more robust assessment of video understanding capabilities. We then propose an architecture that learns spatio-temporal activity localization without any grounding supervision, using only web multimodal video–text data. Finally, we unlock activity-grounding abilities in pretrained multimodal models entirely without training, via lightweight architectural modifications.
Together, these contributions advance image and video understanding without reliance on manual labels. We develop methods for label-free model pretraining and adaptation, reveal dense grounding capabilities without human annotations, and introduce unsupervised tools that improve pretraining data quality and create fairer, more reliable evaluation benchmarks.