Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, Li Yuan. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Recently, Vision Transformers (ViTs) have been broadly explored in visual recognition. With low efficiency in encoding fine-level features, the performance of ViTs is still inferior to the state-of-the-art CNNs when trained from scratch on a midsize dataset like ImageNet. Through experimental analysis, we find it is because of two reasons: 1) the simple tokenization of input images fails to model the important local structure such as edges and lines, leading to low training sample efficiency; 2) the redundant attention backbone design of ViTs leads to limited feature richness for fixed computation budgets and limited training samples. To overcome such limitations, we present a new simple and generic architecture, termed Vision Outlooker (VOLO), which implements a novel outlook attention operation that dynamically conduct the local feature aggregation mechanism in a sliding window manner across the input image. Unlike self-attention that focuses on modeling global dependencies of local features at a coarse level, our outlook attention targets at encoding finer-level features, which is critical for recognition but ignored by self-attention. Outlook attention breaks the bottleneck of self-attention whose computation cost scales quadratically with the input spatial dimension, and thus is much more memory efficient. Compared to our Tokens-To-Token Vision Transformer (T2T-ViT), VOLO can more efficiently encode fine-level features that are essential for high-performance visual recognition. Experiments show that with only 26.6M learnable parameters, VOLO achieves 84.2% top-1 accuracy on ImageNet-1K without using extra training data, 2.7% better than T2T-ViT with a comparable number of parameters. When the model size is scaled up to 296M parameters, its performance can be further improved to 87.1%, setting a new record for ImageNet-1K classification. In addition, we also take the proposed VOLO as pretrained models and report superior performance on downstream tasks, such as semantic segmentation. Code is available at https://github.com/sail-sg/volo .

Despite recent advancements in neural 3D reconstruction, the dependence on dense multi-view captures restricts their broader applicability. In this work, we propose ViewCrafter, a novel method for synthesizing high-fidelity novel views from single or sparse images with the prior of video diffusion model. Our method takes advantage of the powerful generation capabilities of video diffusion model and the coarse 3D clues offered by point-based representation to generate high-quality video frames with significantly improved camera pose control accuracy. To further enlarge the generation range of novel views, we tailored a progressive view synthesis strategy to expand the point cloud and the areas covered by the novel views, which can be further integrated with a camera trajectory planning algorithm to automatically reveal and address occlusions in different scenes. With ViewCrafter, we can facilitate various applications, such as immersive experiences with real-time rendering by efficiently optimizing a 3D-GS representation using the reconstructed 3D points and the generated novel views, and scene-level text-to-3D generation for more imaginative content creation. Extensive experiments on diverse datasets demonstrate the strong generalization capability and superior performance of our method in synthesizing high-fidelity novel views. Our project webpage and code are available at https://drexubery.github.io/ViewCrafter/.

Recently, remarkable progress has been made in scaling up Large Language Models (LLMs) through the use of the sparse Mixture-of-Expert (MoE) layers without significantly increasing computational cost. However, the transition from a pre-trained LLM to a sparse Large Vision-Language Model (LVLM) with MoE remains an open challenge. Directly fine-tuning an LLM to a sparse LVLM often leads to training collapse, characterized by (1) a large modality feature distribution gap and (2) expert load imbalance. This paper proposes a three-stage decoupled weight training process. In the first two stages, the model learns to adapt the LLM to an LVLM. In the third stage, the FFN weights from the second stage are used as lossless initialization for expert weights, effectively constructing a sparse model with a vast number of parameters while maintaining constant computational cost. Through extensive ablation experiments, we derive three empirical guidelines and propose a sparse LVLM termed MoE-LLaVA. MoE-LLaVA is a MoE-based sparse LVLM architecture, which uniquely activates only the top-$k$ experts through routers during deployment, keeping the remaining experts inactive. Extensive experiments demonstrate that MoE-LLaVA outperforms LLaVA-1.5-7B with an average improvement of 4.6 across nine visual understanding benchmarks. Notably, with only 2.2B active parameters, our MoE-LLaVA shows comparable result with LLaVA-1.5-13B (87.0 vs. 85.9) on POPE benchmark. Our work establishes a baseline for sparse LVLMs and provides empirical guidelines for exploring the sparse LVLMs. Our code is available at: https://github.com/PKU-YuanGroup/MoE-LLaVA.

In this paper, we present Vision Permutator, a conceptually simple and data efficient MLP-like architecture for visual recognition. By realizing the importance of the positional information carried by 2D feature representations, unlike recent MLP-like models that encode the spatial information along the flattened spatial dimensions, Vision Permutator separately encodes the feature representations along the height and width dimensions with linear projections. This allows Vision Permutator to capture long-range dependencies and meanwhile avoid the attention building process in transformers. The outputs are then aggregated in a mutually complementing manner to form expressive representations. We show that our Vision Permutators are formidable competitors to convolutional neural networks (CNNs) and vision transformers. Without the dependence on spatial convolutions or attention mechanisms, Vision Permutator achieves 81.5% top-1 accuracy on ImageNet without extra large-scale training data (e.g., ImageNet-22k) using only 25M learnable parameters, which is much better than most CNNs and vision transformers under the same model size constraint. When scaling up to 88M, it attains 83.2% top-1 accuracy, greatly improving the performance of recent state-of-the-art MLP-like networks for visual recognition. We hope this work could encourage research on rethinking the way of encoding spatial information and facilitate the development of MLP-like models. PyTorch/MindSpore/Jittor code is available at https://github.com/Andrew-Qibin/VisionPermutator .