Vision-Language Instruction Tuning: A Review and Analysis

Instruction tuning is a crucial supervised training phase in Large Language Models (LLMs), aiming to enhance the LLM's ability to generalize instruction execution and adapt to user preferences. With the increasing integration of multi-modal data into LLMs, there is growing interest in Vision-Language Instruction Tuning (VLIT), which presents more complex characteristics compared to pure text instruction tuning. In this paper, we systematically review the latest VLIT settings and corresponding datasets in multi-modal LLMs and provide insights into the intrinsic motivations behind their design. For the first time, we offer a detailed multi-perspective categorization for existing VLIT datasets and identify the characteristics that high-quality VLIT data should possess. By incorporating these characteristics as guiding principles into the existing VLIT data construction process, we conduct extensive experiments and verify their positive impact on the performance of tuned multi-modal LLMs. Furthermore, we discuss the current challenges and future research directions of VLIT, providing insights for the continuous development of this field. The code and dataset related to this paper have been open-sourced at https://github.com/palchenli/VL-Instruction-Tuning.

HumTrans: A Novel Open-Source Dataset for Humming Melody Transcription and Beyond

This paper introduces the HumTrans dataset, which is publicly available and primarily designed for humming melody transcription. The dataset can also serve as a foundation for downstream tasks such as humming melody based music generation. It consists of 500 musical compositions of different genres and languages, with each composition divided into multiple segments. In total, the dataset comprises 1000 music segments. To collect this humming dataset, we employed 10 college students, all of whom are either music majors or proficient in playing at least one musical instrument. Each of them hummed every segment twice using the web recording interface provided by our designed website. The humming recordings were sampled at a frequency of 44,100 Hz. During the humming session, the main interface provides a musical score for students to reference, with the melody audio playing simultaneously to aid in capturing both melody and rhythm. The dataset encompasses approximately 56.22 hours of audio, making it the largest known humming dataset to date. The dataset will be released on Hugging Face, and we will provide a GitHub repository containing baseline results and evaluation codes.

TagGPT: Large Language Models are Zero-shot Multimodal Taggers

Tags are pivotal in facilitating the effective distribution of multimedia content in various applications in the contemporary Internet era, such as search engines and recommendation systems. Recently, large language models (LLMs) have demonstrated impressive capabilities across a wide range of tasks. In this work, we propose TagGPT, a fully automated system capable of tag extraction and multimodal tagging in a completely zero-shot fashion. Our core insight is that, through elaborate prompt engineering, LLMs are able to extract and reason about proper tags given textual clues of multimodal data, e.g., OCR, ASR, title, etc. Specifically, to automatically build a high-quality tag set that reflects user intent and interests for a specific application, TagGPT predicts large-scale candidate tags from a series of raw data via prompting LLMs, filtered with frequency and semantics. Given a new entity that needs tagging for distribution, TagGPT introduces two alternative options for zero-shot tagging, i.e., a generative method with late semantic matching with the tag set, and another selective method with early matching in prompts. It is well noticed that TagGPT provides a system-level solution based on a modular framework equipped with a pre-trained LLM (GPT-3.5 used here) and a sentence embedding model (SimCSE used here), which can be seamlessly replaced with any more advanced one you want. TagGPT is applicable for various modalities of data in modern social media and showcases strong generalization ability to a wide range of applications. We evaluate TagGPT on publicly available datasets, i.e., Kuaishou and Food.com, and demonstrate the effectiveness of TagGPT compared to existing hashtags and off-the-shelf taggers. Project page: https://github.com/TencentARC/TagGPT.

Masked Visual Reconstruction in Language Semantic Space

Both masked image modeling (MIM) and natural language supervision have facilitated the progress of transferable visual pre-training. In this work, we seek the synergy between two paradigms and study the emerging properties when MIM meets natural language supervision. To this end, we present a novel masked visual Reconstruction In Language semantic Space (RILS) pre-training framework, in which sentence representations, encoded by the text encoder, serve as prototypes to transform the vision-only signals into patch-sentence probabilities as semantically meaningful MIM reconstruction targets. The vision models can therefore capture useful components with structured information by predicting proper semantic of masked tokens. Better visual representations could, in turn, improve the text encoder via the image-text alignment objective, which is essential for the effective MIM target transformation. Extensive experimental results demonstrate that our method not only enjoys the best of previous MIM and CLIP but also achieves further improvements on various tasks due to their mutual benefits. RILS exhibits advanced transferability on downstream classification, detection, and segmentation, especially for low-shot regimes. Code will be made available at https://github.com/hustvl/RILS.

Masked Image Modeling with Denoising Contrast

Since the development of self-supervised visual representation learning from contrastive learning to masked image modeling, there is no significant difference in essence, that is, how to design proper pretext tasks for vision dictionary look-up. Masked image modeling recently dominates this line of research with state-of-the-art performance on vision Transformers, where the core is to enhance the patch-level visual context capturing of the network via denoising auto-encoding mechanism. Rather than tailoring image tokenizers with extra training stages as in previous works, we unleash the great potential of contrastive learning on denoising auto-encoding and introduce a new pre-training method, ConMIM, to produce simple intra-image inter-patch contrastive constraints as the learning objectives for masked patch prediction. We further strengthen the denoising mechanism with asymmetric designs, including image perturbations and model progress rates, to improve the network pre-training. ConMIM-pretrained vision Transformers with various scales achieve promising results on downstream image classification, semantic segmentation, object detection, and instance segmentation tasks.

Controllable Augmentations for Video Representation Learning

This paper focuses on self-supervised video representation learning. Most existing approaches follow the contrastive learning pipeline to construct positive and negative pairs by sampling different clips. However, this formulation tends to bias to static background and have difficulty establishing global temporal structures. The major reason is that the positive pairs, i.e., different clips sampled from the same video, have limited temporal receptive field, and usually share similar background but differ in motions. To address these problems, we propose a framework to jointly utilize local clips and global videos to learn from detailed region-level correspondence as well as general long-term temporal relations. Based on a set of controllable augmentations, we achieve accurate appearance and motion pattern alignment through soft spatio-temporal region contrast. Our formulation is able to avoid the low-level redundancy shortcut by mutual information minimization to improve the generalization. We also introduce local-global temporal order dependency to further bridge the gap between clip-level and video-level representations for robust temporal modeling. Extensive experiments demonstrate that our framework is superior on three video benchmarks in action recognition and video retrieval, capturing more accurate temporal dynamics.

BridgeFormer: Bridging Video-text Retrieval with Multiple Choice Questions

Pre-training a model to learn transferable video-text representation for retrieval has attracted a lot of attention in recent years. Previous dominant works mainly adopt two separate encoders for efficient retrieval, but ignore local associations between videos and texts. Another line of research uses a joint encoder to interact video with texts, but results in low efficiency since each text-video pair needs to be fed into the model. In this work, we enable fine-grained video-text interactions while maintaining high efficiency for retrieval via a novel pretext task, dubbed as Multiple Choice Questions (MCQ), where a parametric module BridgeFormer is trained to answer the "questions" constructed by the text features via resorting to the video features. Specifically, we exploit the rich semantics of text (i.e., nouns and verbs) to build questions, with which the video encoder can be trained to capture more regional content and temporal dynamics. In the form of questions and answers, the semantic associations between local video-text features can be properly established. BridgeFormer is able to be removed for downstream retrieval, rendering an efficient and flexible model with only two encoders. Our method outperforms state-of-the-art methods on the popular text-to-video retrieval task in five datasets with different experimental setups (i.e., zero-shot and fine-tune), including HowTo100M (one million videos). We further conduct zero-shot action recognition, which can be cast as video-to-text retrieval, and our approach also significantly surpasses its counterparts. As an additional benefit, our method achieves competitive results with much shorter pre-training videos on single-modality downstream tasks, e.g., action recognition with linear evaluation.

CaSP: Class-agnostic Semi-Supervised Pretraining for Detection and Segmentation

To improve instance-level detection/segmentation performance, existing self-supervised and semi-supervised methods extract either very task-unrelated or very task-specific training signals from unlabeled data. We argue that these two approaches, at the two extreme ends of the task-specificity spectrum, are suboptimal for the task performance. Utilizing too little task-specific training signals causes underfitting to the ground-truth labels of downstream tasks, while the opposite causes overfitting to the ground-truth labels. To this end, we propose a novel Class-agnostic Semi-supervised Pretraining (CaSP) framework to achieve a more favorable task-specificity balance in extracting training signals from unlabeled data. Compared to semi-supervised learning, CaSP reduces the task specificity in training signals by ignoring class information in the pseudo labels and having a separate pretraining stage that uses only task-unrelated unlabeled data. On the other hand, CaSP preserves the right amount of task specificity by leveraging box/mask-level pseudo labels. As a result, our pretrained model can better avoid underfitting/overfitting to ground-truth labels when finetuned on the downstream task. Using 3.6M unlabeled data, we achieve a remarkable performance gain of 4.7% over ImageNet-pretrained baseline on object detection. Our pretrained model also demonstrates excellent transferability to other detection and segmentation tasks/frameworks.

CLIP4Caption ++: Multi-CLIP for Video Caption

This report describes our solution to the VALUE Challenge 2021 in the captioning task. Our solution, named CLIP4Caption++, is built on X-Linear/X-Transformer, which is an advanced model with encoder-decoder architecture. We make the following improvements on the proposed CLIP4Caption++: We employ an advanced encoder-decoder model architecture X-Transformer as our main framework and make the following improvements: 1) we utilize three strong pre-trained CLIP models to extract the text-related appearance visual features. 2) we adopt the TSN sampling strategy for data enhancement. 3) we involve the video subtitle information to provide richer semantic information. 3) we introduce the subtitle information, which fuses with the visual features as guidance. 4) we design word-level and sentence-level ensemble strategies. Our proposed method achieves 86.5, 148.4, 64.5 CIDEr scores on VATEX, YC2C, and TVC datasets, respectively, which shows the superior performance of our proposed CLIP4Caption++ on all three datasets.