Large-scale multi-modal contrastive pre-training has demonstrated great utility to learn transferable features for a range of downstream tasks by mapping multiple modalities into a shared embedding space. Typically, this has employed separate encoders for each modality. However, recent work suggests that transformers can support learning across multiple modalities and allow knowledge sharing. Inspired by this, we investigate a variety of Modality-Shared Contrastive Language-Image Pre-training (MS-CLIP) frameworks. More specifically, we question how many parameters of a transformer model can be shared across modalities during contrastive pre-training, and rigorously examine architectural design choices that position the proportion of parameters shared along a spectrum. In studied conditions, we observe that a mostly unified encoder for vision and language signals outperforms all other variations that separate more parameters. Additionally, we find that light-weight modality-specific parallel modules further improve performance. Experimental results show that the proposed MS-CLIP approach outperforms vanilla CLIP by up to 13\% relative in zero-shot ImageNet classification (pre-trained on YFCC-100M), while simultaneously supporting a reduction of parameters. In addition, our approach outperforms vanilla CLIP by 1.6 points in linear probing on a collection of 24 downstream vision tasks. Furthermore, we discover that sharing parameters leads to semantic concepts from different modalities being encoded more closely in the embedding space, facilitating the transferring of common semantic structure (e.g., attention patterns) from language to vision. Code is available at \href{https://github.com/Hxyou/MSCLIP}{URL}.
Vision transformer (ViT) recently has drawn great attention in computer vision due to its remarkable model capability. However, most prevailing ViT models suffer from huge number of parameters, restricting their applicability on devices with limited resources. To alleviate this issue, we propose TinyViT, a new family of tiny and efficient small vision transformers pretrained on large-scale datasets with our proposed fast distillation framework. The central idea is to transfer knowledge from large pretrained models to small ones, while enabling small models to get the dividends of massive pretraining data. More specifically, we apply distillation during pretraining for knowledge transfer. The logits of large teacher models are sparsified and stored in disk in advance to save the memory cost and computation overheads. The tiny student transformers are automatically scaled down from a large pretrained model with computation and parameter constraints. Comprehensive experiments demonstrate the efficacy of TinyViT. It achieves a top-1 accuracy of 84.8% on ImageNet-1k with only 21M parameters, being comparable to Swin-B pretrained on ImageNet-21k while using 4.2 times fewer parameters. Moreover, increasing image resolutions, TinyViT can reach 86.5% accuracy, being slightly better than Swin-L while using only 11% parameters. Last but not the least, we demonstrate a good transfer ability of TinyViT on various downstream tasks. Code and models are available at https://github.com/microsoft/Cream/tree/main/TinyViT.
We propose bootstrapped masked autoencoders (BootMAE), a new approach for vision BERT pretraining. BootMAE improves the original masked autoencoders (MAE) with two core designs: 1) momentum encoder that provides online feature as extra BERT prediction targets; 2) target-aware decoder that tries to reduce the pressure on the encoder to memorize target-specific information in BERT pretraining. The first design is motivated by the observation that using a pretrained MAE to extract the features as the BERT prediction target for masked tokens can achieve better pretraining performance. Therefore, we add a momentum encoder in parallel with the original MAE encoder, which bootstraps the pretraining performance by using its own representation as the BERT prediction target. In the second design, we introduce target-specific information (e.g., pixel values of unmasked patches) from the encoder directly to the decoder to reduce the pressure on the encoder of memorizing the target-specific information. Thus, the encoder focuses on semantic modeling, which is the goal of BERT pretraining, and does not need to waste its capacity in memorizing the information of unmasked tokens related to the prediction target. Through extensive experiments, our BootMAE achieves $84.2\%$ Top-1 accuracy on ImageNet-1K with ViT-B backbone, outperforming MAE by $+0.8\%$ under the same pre-training epochs. BootMAE also gets $+1.0$ mIoU improvements on semantic segmentation on ADE20K and $+1.3$ box AP, $+1.4$ mask AP improvement on object detection and segmentation on COCO dataset. Code is released at https://github.com/LightDXY/BootMAE.
The complexity-precision trade-off of an object detector is a critical problem for resource constrained vision tasks. Previous works have emphasized detectors implemented with efficient backbones. The impact on this trade-off of proposal processing by the detection head is investigated in this work. It is hypothesized that improved detection efficiency requires a paradigm shift, towards the unequal processing of proposals, assigning more computation to good proposals than poor ones. This results in better utilization of available computational budget, enabling higher accuracy for the same FLOPS. We formulate this as a learning problem where the goal is to assign operators to proposals, in the detection head, so that the total computational cost is constrained and the precision is maximized. The key finding is that such matching can be learned as a function that maps each proposal embedding into a one-hot code over operators. While this function induces a complex dynamic network routing mechanism, it can be implemented by a simple MLP and learned end-to-end with off-the-shelf object detectors. This 'dynamic proposal processing' (DPP) is shown to outperform state-of-the-art end-to-end object detectors (DETR, Sparse R-CNN) by a clear margin for a given computational complexity.
Denoising Diffusion Probabilistic Models (DDPMs) have achieved remarkable success in various image generation tasks compared with Generative Adversarial Nets (GANs). Recent work on semantic image synthesis mainly follows the \emph{de facto} GAN-based approaches, which may lead to unsatisfactory quality or diversity of generated images. In this paper, we propose a novel framework based on DDPM for semantic image synthesis. Unlike previous conditional diffusion model directly feeds the semantic layout and noisy image as input to a U-Net structure, which may not fully leverage the information in the input semantic mask, our framework processes semantic layout and noisy image differently. It feeds noisy image to the encoder of the U-Net structure while the semantic layout to the decoder by multi-layer spatially-adaptive normalization operators. To further improve the generation quality and semantic interpretability in semantic image synthesis, we introduce the classifier-free guidance sampling strategy, which acknowledge the scores of an unconditional model for sampling process. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our proposed method, achieving state-of-the-art performance in terms of fidelity~(FID) and diversity~(LPIPS).
We present GLIPv2, a grounded VL understanding model, that serves both localization tasks (e.g., object detection, instance segmentation) and Vision-Language (VL) understanding tasks (e.g., VQA, image captioning). GLIPv2 elegantly unifies localization pre-training and Vision-Language Pre-training (VLP) with three pre-training tasks: phrase grounding as a VL reformulation of the detection task, region-word contrastive learning as a novel region-word level contrastive learning task, and the masked language modeling. This unification not only simplifies the previous multi-stage VLP procedure but also achieves mutual benefits between localization and understanding tasks. Experimental results show that a single GLIPv2 model (all model weights are shared) achieves near SoTA performance on various localization and understanding tasks. The model also shows (1) strong zero-shot and few-shot adaption performance on open-vocabulary object detection tasks and (2) superior grounding capability on VL understanding tasks. Code will be released at https://github.com/microsoft/GLIP.
Leveraging large-scale data can introduce performance gains on many computer vision tasks. Unfortunately, this does not happen in object detection when training a single model under multiple datasets together. We observe two main obstacles: taxonomy difference and bounding box annotation inconsistency, which introduces domain gaps in different datasets that prevents us from joint training. In this paper, we show that these two challenges can be effectively addressed by simply adapting object queries on language embedding of categories per dataset. We design a detection hub to dynamically adapt queries on category embedding based on the different distributions of datasets. Unlike previous methods attempted to learn a joint embedding for all datasets, our adaptation method can utilize the language embedding as semantic centers for common categories, while learning the semantic bias towards specific categories belonging to different datasets to handle annotation differences and make up the domain gaps. These novel improvements enable us to end-to-end train a single detector on multiple datasets simultaneously to fully take their advantages. Further experiments on joint training on multiple datasets demonstrate the significant performance gains over separate individual fine-tuned detectors.
People say, "A picture is worth a thousand words". Then how can we get the rich information out of the image? We argue that by using visual clues to bridge large pretrained vision foundation models and language models, we can do so without any extra cross-modal training. Thanks to the strong zero-shot capability of foundation models, we start by constructing a rich semantic representation of the image (e.g., image tags, object attributes / locations, captions) as a structured textual prompt, called visual clues, using a vision foundation model. Based on visual clues, we use large language model to produce a series of comprehensive descriptions for the visual content, which is then verified by the vision model again to select the candidate that aligns best with the image. We evaluate the quality of generated descriptions by quantitative and qualitative measurement. The results demonstrate the effectiveness of such a structured semantic representation.
This paper revisits visual representation in knowledge-based visual question answering (VQA) and demonstrates that using regional information in a better way can significantly improve the performance. While visual representation is extensively studied in traditional VQA, it is under-explored in knowledge-based VQA even though these two tasks share the common spirit, i.e., rely on visual input to answer the question. Specifically, we observe that in most state-of-the-art knowledge-based VQA methods: 1) visual features are extracted either from the whole image or in a sliding window manner for retrieving knowledge, and the important relationship within/among object regions is neglected; 2) visual features are not well utilized in the final answering model, which is counter-intuitive to some extent. Based on these observations, we propose a new knowledge-based VQA method REVIVE, which tries to utilize the explicit information of object regions not only in the knowledge retrieval stage but also in the answering model. The key motivation is that object regions and inherent relationships are important for knowledge-based VQA. We perform extensive experiments on the standard OK-VQA dataset and achieve new state-of-the-art performance, i.e., 58.0% accuracy, surpassing previous state-of-the-art method by a large margin (+3.6%). We also conduct detailed analysis and show the necessity of regional information in different framework components for knowledge-based VQA.
Transformers have achieved great success in pluralistic image inpainting recently. However, we find existing transformer based solutions regard each pixel as a token, thus suffer from information loss issue from two aspects: 1) They downsample the input image into much lower resolutions for efficiency consideration, incurring information loss and extra misalignment for the boundaries of masked regions. 2) They quantize $256^3$ RGB pixels to a small number (such as 512) of quantized pixels. The indices of quantized pixels are used as tokens for the inputs and prediction targets of transformer. Although an extra CNN network is used to upsample and refine the low-resolution results, it is difficult to retrieve the lost information back.To keep input information as much as possible, we propose a new transformer based framework "PUT". Specifically, to avoid input downsampling while maintaining the computation efficiency, we design a patch-based auto-encoder P-VQVAE, where the encoder converts the masked image into non-overlapped patch tokens and the decoder recovers the masked regions from inpainted tokens while keeping the unmasked regions unchanged. To eliminate the information loss caused by quantization, an Un-Quantized Transformer (UQ-Transformer) is applied, which directly takes the features from P-VQVAE encoder as input without quantization and regards the quantized tokens only as prediction targets. Extensive experiments show that PUT greatly outperforms state-of-the-art methods on image fidelity, especially for large masked regions and complex large-scale datasets. Code is available at https://github.com/liuqk3/PUT