Large Multimodal Model (LMM) is a hot research topic in the computer vision area and has also demonstrated remarkable potential across multiple disciplinary fields. A recent trend is to further extend and enhance the perception capabilities of LMMs. The current methods follow the paradigm of adapting the visual task outputs to the format of the language model, which is the main component of a LMM. This adaptation leads to convenient development of such LMMs with minimal modifications, however, it overlooks the intrinsic characteristics of diverse visual tasks and hinders the learning of perception capabilities. To address this issue, we propose a novel LMM architecture named Lumen, a Large multimodal model with versatile vision-centric capability enhancement. We decouple the LMM's learning of perception capabilities into task-agnostic and task-specific stages. Lumen first promotes fine-grained vision-language concept alignment, which is the fundamental capability for various visual tasks. Thus the output of the task-agnostic stage is a shared representation for all the tasks we address in this paper. Then the task-specific decoding is carried out by flexibly routing the shared representation to lightweight task decoders with negligible training efforts. Benefiting from such a decoupled design, our Lumen surpasses existing LMM-based approaches on the COCO detection benchmark with a clear margin and exhibits seamless scalability to additional visual tasks. Furthermore, we also conduct comprehensive ablation studies and generalization evaluations for deeper insights. The code will be released at https://github.com/SxJyJay/Lumen.
Instruction finetuning on a variety of image-text instruction data is the key to obtaining a versatile Multimodal Large Language Model (MLLM), and different configurations of the instruction data can lead to finetuned models with different capabilities. However, we have discovered that data conflicts are inevitable when mixing instruction data from distinct domains, which can result in performance drops for tasks of a specific domain. To address this issue, we propose to apply an efficient Mixture of Experts (MoE) design, which is a sparse Mixture of LoRA Experts (MoLE) for instruction finetuning MLLMs. Within the Transformer layers, we extend the popular Low-Rank Adaption (LoRA) method by creating a set of LoRA experts specifically for the MLP layer, and route each token to the top-1 expert based on a routing function, allowing adaptive choices for tokens from different domains. Since the LoRA experts are sparsely activated, the training and inference cost are kept roughly constant compared to the original LoRA method. By replacing the plain-LoRA of LLaVA-1.5 with our MoE design, our final model is named LLaVA-MoLE. Extensive experiments proved that LLaVA-MoLE effectively mitigates the data conflict issue when mixing multiple distinct instruction datasets with various configurations, and achieves consistent performance gains over the strong plain-LoRA baselines. Most importantly, on the mixed datasets, LLaVA-MoLE can even outperform the plain-LoRA baseline trained with twice the samples.
Camera-based bird-eye-view (BEV) perception paradigm has made significant progress in the autonomous driving field. Under such a paradigm, accurate BEV representation construction relies on reliable depth estimation for multi-camera images. However, existing approaches exhaustively predict depths for every pixel without prioritizing objects, which are precisely the entities requiring detection in the 3D space. To this end, we propose IA-BEV, which integrates image-plane instance awareness into the depth estimation process within a BEV-based detector. First, a category-specific structural priors mining approach is proposed for enhancing the efficacy of monocular depth generation. Besides, a self-boosting learning strategy is further proposed to encourage the model to place more emphasis on challenging objects in computation-expensive temporal stereo matching. Together they provide advanced depth estimation results for high-quality BEV features construction, benefiting the ultimate 3D detection. The proposed method achieves state-of-the-art performances on the challenging nuScenes benchmark, and extensive experimental results demonstrate the effectiveness of our designs.
The task of moment localization is to localize a temporal moment in an untrimmed video for a given natural language query. Since untrimmed video contains highly redundant contents, the quality of the query is crucial for accurately localizing moments, i.e., the query should provide precise information about the target moment so that the localization model can understand what to look for in the videos. However, the natural language queries in current datasets may not be easy to understand for existing models. For example, the Ego4D dataset uses question sentences as the query to describe relatively complex moments. While being natural and straightforward for humans, understanding such question sentences are challenging for mainstream moment localization models like 2D-TAN. Inspired by the recent success of large language models, especially their ability of understanding and generating complex natural language contents, in this extended abstract, we make early attempts at reformulating the moment queries into a set of instructions using large language models and making them more friendly to the localization models.
Fusing LiDAR and camera information is essential for achieving accurate and reliable 3D object detection in autonomous driving systems. However, this is challenging due to the difficulty of combining multi-granularity geometric and semantic features from two drastically different modalities. Recent approaches aim at exploring the semantic densities of camera features through lifting points in 2D camera images (referred to as seeds) into 3D space for fusion, and they can be roughly divided into 1) early fusion of raw points that aims at augmenting the 3D point cloud at the early input stage, and 2) late fusion of BEV (bird-eye view) maps that merges LiDAR and camera BEV features before the detection head. While both have their merits in enhancing the representation power of the combined features, this single-level fusion strategy is a suboptimal solution to the aforementioned challenge. Their major drawbacks are the inability to interact the multi-granularity semantic features from two distinct modalities sufficiently. To this end, we propose a novel framework that focuses on the multi-scale progressive interaction of the multi-granularity LiDAR and camera features. Our proposed method, abbreviated as MDMSFusion, achieves state-of-the-art results in 3D object detection, with 69.1 mAP and 71.8 NDS on nuScenes validation set, and 70.8 mAP and 73.2 NDS on nuScenes test set, which rank 1st and 2nd respectively among single-model non-ensemble approaches by the time of submission.
In this technical report, we introduce our submission to the Waymo 3D Detection leaderboard. Our network is based on the Centerpoint architecture, but with significant improvements. We design a 2D backbone to utilize multi-scale features for better detecting objects with various sizes, together with an optimal transport-based target assignment strategy, which dynamically assigns richer supervision signals to the detection candidates. We also apply test-time augmentation and model-ensemble for further improvements. Our submission currently ranks 4th place with 78.45 mAPH on the Waymo 3D Detection leaderboard.
3D dense captioning is a recently-proposed novel task, where point clouds contain more geometric information than the 2D counterpart. However, it is also more challenging due to the higher complexity and wider variety of inter-object relations. Existing methods only treat such relations as by-products of object feature learning in graphs without specifically encoding them, which leads to sub-optimal results. In this paper, aiming at improving 3D dense captioning via capturing and utilizing the complex relations in the 3D scene, we propose MORE, a Multi-Order RElation mining model, to support generating more descriptive and comprehensive captions. Technically, our MORE encodes object relations in a progressive manner since complex relations can be deduced from a limited number of basic ones. We first devise a novel Spatial Layout Graph Convolution (SLGC), which semantically encodes several first-order relations as edges of a graph constructed over 3D object proposals. Next, from the resulting graph, we further extract multiple triplets which encapsulate basic first-order relations as the basic unit and construct several Object-centric Triplet Attention Graphs (OTAG) to infer multi-order relations for every target object. The updated node features from OTAG are aggregated and fed into the caption decoder to provide abundant relational cues so that captions including diverse relations with context objects can be generated. Extensive experiments on the Scan2Cap dataset prove the effectiveness of our proposed MORE and its components, and we also outperform the current state-of-the-art method.
Given a text description, Temporal Language Grounding (TLG) aims to localize temporal boundaries of the segments that contain the specified semantics in an untrimmed video. TLG is inherently a challenging task, as it requires to have comprehensive understanding of both video contents and text sentences. Previous works either tackle this task in a fully-supervised setting that requires a large amount of manual annotations or in a weakly supervised setting that cannot achieve satisfactory performance. To achieve good performance with limited annotations, we tackle this task in a semi-supervised way and propose a unified Semi-supervised Temporal Language Grounding (STLG) framework. STLG consists of two parts: (1) A pseudo label generation module that produces adaptive instant pseudo labels for unlabeled data based on predictions from a teacher model; (2) A self-supervised feature learning module with two sequential perturbations, i.e., time lagging and time scaling, for improving the video representation by inter-modal and intra-modal contrastive learning. We conduct experiments on the ActivityNet-CD-OOD and Charades-CD-OOD datasets and the results demonstrate that our proposed STLG framework achieve competitive performance compared to fully-supervised state-of-the-art methods with only a small portion of temporal annotations.
Blind face inpainting refers to the task of reconstructing visual contents without explicitly indicating the corrupted regions in a face image. Inherently, this task faces two challenges: (1) how to detect various mask patterns of different shapes and contents; (2) how to restore visually plausible and pleasing contents in the masked regions. In this paper, we propose a novel two-stage blind face inpainting method named Frequency-guided Transformer and Top-Down Refinement Network (FT-TDR) to tackle these challenges. Specifically, we first use a transformer-based network to detect the corrupted regions to be inpainted as masks by modeling the relation among different patches. We also exploit the frequency modality as complementary information for improved detection results and capture the local contextual incoherence to enhance boundary consistency. Then a top-down refinement network is proposed to hierarchically restore features at different levels and generate contents that are semantically consistent with the unmasked face regions. Extensive experiments demonstrate that our method outperforms current state-of-the-art blind and non-blind face inpainting methods qualitatively and quantitatively.
Automatically generating sentences to describe events and temporally localizing sentences in a video are two important tasks that bridge language and videos. Recent techniques leverage the multimodal nature of videos by using off-the-shelf features to represent videos, but interactions between modalities are rarely explored. Inspired by the fact that there exist cross-modal interactions in the human brain, we propose a novel method for learning pairwise modality interactions in order to better exploit complementary information for each pair of modalities in videos and thus improve performances on both tasks. We model modality interaction in both the sequence and channel levels in a pairwise fashion, and the pairwise interaction also provides some explainability for the predictions of target tasks. We demonstrate the effectiveness of our method and validate specific design choices through extensive ablation studies. Our method turns out to achieve state-of-the-art performances on four standard benchmark datasets: MSVD and MSR-VTT (event captioning task), and Charades-STA and ActivityNet Captions (temporal sentence localization task).