Fueled by the Large Language Models (LLMs) wave, Large Visual-Language Models (LVLMs) have emerged as a pivotal advancement, bridging the gap between image and text. However, video making it challenging for LVLMs to perform adequately due to the complexity of the relationship between language and spatial-temporal data structure. Recent Large Video-Language Models (LVidLMs) align feature of static visual data like image into latent space of language feature, by general multi-modal tasks to leverage abilities of LLMs sufficiently. In this paper, we explore fine-grained alignment approach via object trajectory for different modalities across both spatial and temporal dimensions simultaneously. Thus, we propose a novel LVidLM by trajectory-guided Pixel-Temporal Alignment, dubbed PiTe, that exhibits promising applicable model property. To achieve fine-grained video-language alignment, we curate a multi-modal pre-training dataset PiTe-143k, the dataset provision of moving trajectories in pixel level for all individual objects, that appear and mention in the video and caption both, by our automatic annotation pipeline. Meanwhile, PiTe demonstrates astounding capabilities on myriad video-related multi-modal tasks through beat the state-of-the-art methods by a large margin.

Self-supervised learning (SSL) has proven effective in learning high-quality representations for various downstream tasks, with a primary focus on semantic tasks. However, its application in geometric tasks remains underexplored, partially due to the absence of a standardized evaluation method for geometric representations. To address this gap, we introduce a novel pose-estimation benchmark for assessing SSL geometric representations, which demands training without semantic or pose labels and achieving proficiency in both semantic and geometric downstream tasks. On this benchmark, we study enhancing SSL geometric representations without sacrificing semantic classification accuracy. We find that leveraging mid-layer representations improves pose-estimation performance by 10-20%. Further, we introduce an unsupervised trajectory-regularization loss, which improves performance by an additional 4% and improves generalization ability on out-of-distribution data. We hope the proposed benchmark and methods offer new insights and improvements in self-supervised geometric representation learning.

While recent vision-and-language models (VLMs) like CLIP are a powerful tool for analyzing text and images in a shared semantic space, they do not explicitly model the hierarchical nature of the set of texts which may describe an image. Conversely, existing multimodal hierarchical representation learning methods require costly training from scratch, failing to leverage the knowledge encoded by state-of-the-art multimodal foundation models. In this work, we study the knowledge of existing foundation models, finding that they exhibit emergent understanding of visual-semantic hierarchies despite not being directly trained for this purpose. We propose the Radial Embedding (RE) framework for probing and optimizing hierarchical understanding, and contribute the HierarCaps dataset, a benchmark facilitating the study of hierarchical knowledge in image--text representations, constructed automatically via large language models. Our results show that foundation VLMs exhibit zero-shot hierarchical understanding, surpassing the performance of prior models explicitly designed for this purpose. Furthermore, we show that foundation models may be better aligned to hierarchical reasoning via a text-only fine-tuning phase, while retaining pretraining knowledge. We will release our data, code, and trained models.

Multimodal models have gained increasing popularity recently. Many works have been proposed to learn the representations for different modalities. The representation can learn shared information from these domains, leading to increased and coherent joint and cross-generation. However, these works mainly considered standard Gaussian or Laplacian as their prior distribution. It can be challenging for the uni-modal and non-informative distribution to capture all the information from multiple data types. Meanwhile, energy-based models (EBM) have shown their effectiveness in multiple tasks due to their expressiveness and flexibility. But its capacity has yet to be discovered for the multimodal generative models. In this paper, we propose a novel framework to train multimodal latent generative models together with the energy-based models. The proposed method can lead to more expressive and informative prior which can better capture the information within multiple modalities. Our experiments showed that our model is effective and can increase generation coherence and latent classification for different multimodal datasets.

The increasing availability of multi-sensor data sparks interest in multimodal self-supervised learning. However, most existing approaches learn only common representations across modalities while ignoring intra-modal training and modality-unique representations. We propose Decoupling Common and Unique Representations (DeCUR), a simple yet effective method for multimodal self-supervised learning. By distinguishing inter- and intra-modal embeddings through multimodal redundancy reduction, DeCUR can integrate complementary information across different modalities. Meanwhile, a simple residual deformable attention is introduced to help the model focus on modality-informative features. We evaluate DeCUR in three common multimodal scenarios ( radar-optical, RGB-elevation, and RGB-depth), and demonstrate its consistent and significant improvement for both multimodal and modality-missing settings. With thorough experiments and comprehensive analysis, we hope this work can provide insights and raise more interest in researching the hidden relationships of multimodal representations.

Vision Transformer models trained on large-scale datasets, although effective, often exhibit artifacts in the patch token they extract. While such defects can be alleviated by re-training the entire model with additional classification tokens, the underlying reasons for the presence of these tokens remain unclear. In this paper, we conduct a thorough investigation of this phenomenon, combining theoretical analysis with empirical observations. Our findings reveal that these artifacts originate from the pre-trained network itself, specifically stemming from the leading left singular vector of the network's weights. Furthermore, to mitigate these defects, we propose a novel fine-tuning smooth regularization that rectifies structural deficiencies using only a small dataset, thereby avoiding the need for complete re-training. We validate our method on various downstream tasks, including unsupervised segmentation, classification, and supervised segmentation, demonstrating its effectiveness in improving model performance. Our code and checkpoints will be released.

We delve into a crucial yet often overlooked challenge inherent to Vision Transformers (ViTs): feature maps of these models exhibit grid-like artifacts, which hurt the performance of ViTs in downstream dense prediction tasks such as segmentation, depth prediction, and object discovery. We trace this fundamental issue down to the positional embeddings at the input stage. we propose a two-stage denoising approach, termed Denoising Vision Transformers (DVT). In the first stage, we separate the clean features from those contaminated by positional artifacts by enforcing cross-view feature consistency with neural fields on a per-image basis. This per-image optimization process extracts artifact-free features from raw ViT outputs, providing clean feature estimates for offline applications. In the second stage, we train a lightweight Transformer block to predict clean features from raw ViT outputs, leveraging the derived estimates of the clean features as supervision. Our DVT does not require re-training the existing pre-trained ViTs, and is immediately applicable to any Vision Transformer architecture. We evaluate our method on a variety of representative ViTs (DINO, DeiT-III, EVA02, CLIP, DINOv2, DINOv2-reg) and demonstrate that our DVT consistently and significantly improves existing state-of-the-art general-purpose models in semantic and geometric tasks across multiple datasets. We hope our study will encourage a re-evaluation of ViT design, especially regarding the naive use of positional embeddings. Our code and models will be released.

Recent advancements in transformer-based light-weight object tracking have set new standards across various benchmarks due to their efficiency and effectiveness. Despite these achievements, most current trackers rely heavily on pre-existing object detection architectures without optimizing the backbone network to leverage the unique demands of object tracking. Addressing this gap, we introduce the Feature Extraction and Relation Modeling Tracker (FERMT) - a novel approach that significantly enhances tracking speed and accuracy. At the heart of FERMT is a strategic decomposition of the conventional attention mechanism into four distinct sub-modules within a one-stream tracker. This design stems from our insight that the initial layers of a tracking network should prioritize feature extraction, whereas the deeper layers should focus on relation modeling between objects. Consequently, we propose an innovative, light-weight backbone specifically tailored for object tracking. Our approach is validated through meticulous ablation studies, confirming the effectiveness of our architectural decisions. Furthermore, FERMT incorporates a Dual Attention Unit for feature pre-processing, which facilitates global feature interaction across channels and enriches feature representation with attention cues. Benchmarking on GOT-10k, FERMT achieves a groundbreaking Average Overlap (AO) score of 69.6%, outperforming the leading real-time trackers by 5.6% in accuracy while boasting a 54% improvement in CPU tracking speed. This work not only sets a new standard for state-of-the-art (SOTA) performance in light-weight tracking but also bridges the efficiency gap between fast and high-performance trackers.The code and model will be available soon.