Yehao Li

Ting Yao, Yehao Li, Yingwei Pan et al.

Prior works have proposed several strategies to reduce the computational cost of self-attention mechanism. Many of these works consider decomposing the self-attention procedure into regional and local feature extraction procedures that each incurs a much smaller computational complexity. However, regional information is typically only achieved at the expense of undesirable information lost owing to down-sampling. In this paper, we propose a novel Transformer architecture that aims to mitigate the cost issue, named Dual Vision Transformer (Dual-ViT). The new architecture incorporates a critical semantic pathway that can more efficiently compress token vectors into global semantics with reduced order of complexity. Such compressed global semantics then serve as useful prior information in learning finer pixel level details, through another constructed pixel pathway. The semantic pathway and pixel pathway are then integrated together and are jointly trained, spreading the enhanced self-attention information in parallel through both of the pathways. Dual-ViT is henceforth able to reduce the computational complexity without compromising much accuracy. We empirically demonstrate that Dual-ViT provides superior accuracy than SOTA Transformer architectures with reduced training complexity. Source code is available at \url{https://github.com/YehLi/ImageNetModel}.

Ting Yao, Yingwei Pan, Yehao Li et al.

Multi-scale Vision Transformer (ViT) has emerged as a powerful backbone for computer vision tasks, while the self-attention computation in Transformer scales quadratically w.r.t. the input patch number. Thus, existing solutions commonly employ down-sampling operations (e.g., average pooling) over keys/values to dramatically reduce the computational cost. In this work, we argue that such over-aggressive down-sampling design is not invertible and inevitably causes information dropping especially for high-frequency components in objects (e.g., texture details). Motivated by the wavelet theory, we construct a new Wavelet Vision Transformer (\textbf{Wave-ViT}) that formulates the invertible down-sampling with wavelet transforms and self-attention learning in a unified way. This proposal enables self-attention learning with lossless down-sampling over keys/values, facilitating the pursuing of a better efficiency-vs-accuracy trade-off. Furthermore, inverse wavelet transforms are leveraged to strengthen self-attention outputs by aggregating local contexts with enlarged receptive field. We validate the superiority of Wave-ViT through extensive experiments over multiple vision tasks (e.g., image recognition, object detection and instance segmentation). Its performances surpass state-of-the-art ViT backbones with comparable FLOPs. Source code is available at \url{https://github.com/YehLi/ImageNetModel}.

Jianjie Luo, Yehao Li, Yingwei Pan et al.

Recent advances on text-to-image generation have witnessed the rise of diffusion models which act as powerful generative models. Nevertheless, it is not trivial to exploit such latent variable models to capture the dependency among discrete words and meanwhile pursue complex visual-language alignment in image captioning. In this paper, we break the deeply rooted conventions in learning Transformer-based encoder-decoder, and propose a new diffusion model based paradigm tailored for image captioning, namely Semantic-Conditional Diffusion Networks (SCD-Net). Technically, for each input image, we first search the semantically relevant sentences via cross-modal retrieval model to convey the comprehensive semantic information. The rich semantics are further regarded as semantic prior to trigger the learning of Diffusion Transformer, which produces the output sentence in a diffusion process. In SCD-Net, multiple Diffusion Transformer structures are stacked to progressively strengthen the output sentence with better visional-language alignment and linguistical coherence in a cascaded manner. Furthermore, to stabilize the diffusion process, a new self-critical sequence training strategy is designed to guide the learning of SCD-Net with the knowledge of a standard autoregressive Transformer model. Extensive experiments on COCO dataset demonstrate the promising potential of using diffusion models in the challenging image captioning task. Source code is available at \url{https://github.com/YehLi/xmodaler/tree/master/configs/image_caption/scdnet}.

Yehao Li, Yingwei Pan, Ting Yao et al.

Comprehending the rich semantics in an image and ordering them in linguistic order are essential to compose a visually-grounded and linguistically coherent description for image captioning. Modern techniques commonly capitalize on a pre-trained object detector/classifier to mine the semantics in an image, while leaving the inherent linguistic ordering of semantics under-exploited. In this paper, we propose a new recipe of Transformer-style structure, namely Comprehending and Ordering Semantics Networks (COS-Net), that novelly unifies an enriched semantic comprehending and a learnable semantic ordering processes into a single architecture. Technically, we initially utilize a cross-modal retrieval model to search the relevant sentences of each image, and all words in the searched sentences are taken as primary semantic cues. Next, a novel semantic comprehender is devised to filter out the irrelevant semantic words in primary semantic cues, and meanwhile infer the missing relevant semantic words visually grounded in the image. After that, we feed all the screened and enriched semantic words into a semantic ranker, which learns to allocate all semantic words in linguistic order as humans. Such sequence of ordered semantic words are further integrated with visual tokens of images to trigger sentence generation. Empirical evidences show that COS-Net clearly surpasses the state-of-the-art approaches on COCO and achieves to-date the best CIDEr score of 141.1% on Karpathy test split. Source code is available at \url{https://github.com/YehLi/xmodaler/tree/master/configs/image_caption/cosnet}.

Zhaofan Qiu, Yehao Li, Yu Wang et al.

In this paper, we propose a novel deep architecture tailored for 3D point cloud applications, named as SPE-Net. The embedded ``Selective Position Encoding (SPE)'' procedure relies on an attention mechanism that can effectively attend to the underlying rotation condition of the input. Such encoded rotation condition then determines which part of the network parameters to be focused on, and is shown to efficiently help reduce the degree of freedom of the optimization during training. This mechanism henceforth can better leverage the rotation augmentations through reduced training difficulties, making SPE-Net robust against rotated data both during training and testing. The new findings in our paper also urge us to rethink the relationship between the extracted rotation information and the actual test accuracy. Intriguingly, we reveal evidences that by locally encoding the rotation information through SPE-Net, the rotation-invariant features are still of critical importance in benefiting the test samples without any actual global rotation. We empirically demonstrate the merits of the SPE-Net and the associated hypothesis on four benchmarks, showing evident improvements on both rotated and unrotated test data over SOTA methods. Source code is available at https://github.com/ZhaofanQiu/SPE-Net.

Siqi Wan, Yehao Li, Jingwen Chen et al.

Diffusion models have led to the revolutionizing of generative modeling in numerous image synthesis tasks. Nevertheless, it is not trivial to directly apply diffusion models for synthesizing an image of a target person wearing a given in-shop garment, i.e., image-based virtual try-on (VTON) task. The difficulty originates from the aspect that the diffusion process should not only produce holistically high-fidelity photorealistic image of the target person, but also locally preserve every appearance and texture detail of the given garment. To address this, we shape a new Diffusion model, namely GarDiff, which triggers the garment-focused diffusion process with amplified guidance of both basic visual appearance and detailed textures (i.e., high-frequency details) derived from the given garment. GarDiff first remoulds a pre-trained latent diffusion model with additional appearance priors derived from the CLIP and VAE encodings of the reference garment. Meanwhile, a novel garment-focused adapter is integrated into the UNet of diffusion model, pursuing local fine-grained alignment with the visual appearance of reference garment and human pose. We specifically design an appearance loss over the synthesized garment to enhance the crucial, high-frequency details. Extensive experiments on VITON-HD and DressCode datasets demonstrate the superiority of our GarDiff when compared to state-of-the-art VTON approaches. Code is publicly available at: \href{https://github.com/siqi0905/GarDiff/tree/master}{https://github.com/siqi0905/GarDiff/tree/master}.

Yifu Chen, Jingwen Chen, Yingwei Pan et al.

Recent years have witnessed the success of large text-to-image diffusion models and their remarkable potential to generate high-quality images. The further pursuit of enhancing the editability of images has sparked significant interest in the downstream task of inpainting a novel object described by a text prompt within a designated region in the image. Nevertheless, the problem is not trivial from two aspects: 1) Solely relying on one single U-Net to align text prompt and visual object across all the denoising timesteps is insufficient to generate desired objects; 2) The controllability of object generation is not guaranteed in the intricate sampling space of diffusion model. In this paper, we propose to decompose the typical single-stage object inpainting into two cascaded processes: 1) semantic pre-inpainting that infers the semantic features of desired objects in a multi-modal feature space; 2) high-fieldity object generation in diffusion latent space that pivots on such inpainted semantic features. To achieve this, we cascade a Transformer-based semantic inpainter and an object inpainting diffusion model, leading to a novel CAscaded Transformer-Diffusion (CAT-Diffusion) framework for text-guided object inpainting. Technically, the semantic inpainter is trained to predict the semantic features of the target object conditioning on unmasked context and text prompt. The outputs of the semantic inpainter then act as the informative visual prompts to guide high-fieldity object generation through a reference adapter layer, leading to controllable object inpainting. Extensive evaluations on OpenImages-V6 and MSCOCO validate the superiority of CAT-Diffusion against the state-of-the-art methods. Code is available at \url{https://github.com/Nnn-s/CATdiffusion}.

Yingwei Pan, Yehao Li, Yiheng Zhang et al.

This paper presents an overview and comparative analysis of our systems designed for the following two tracks in SAPIEN ManiSkill Challenge 2021: No Interaction Track: The No Interaction track targets for learning policies from pre-collected demonstration trajectories. We investigate both imitation learning-based approach, i.e., imitating the observed behavior using classical supervised learning techniques, and offline reinforcement learning-based approaches, for this track. Moreover, the geometry and texture structures of objects and robotic arms are exploited via Transformer-based networks to facilitate imitation learning. No Restriction Track: In this track, we design a Heuristic Rule-based Method (HRM) to trigger high-quality object manipulation by decomposing the task into a series of sub-tasks. For each sub-task, the simple rule-based controlling strategies are adopted to predict actions that can be applied to robotic arms. To ease the implementations of our systems, all the source codes and pre-trained models are available at \url{https://github.com/caiqi/Silver-Bullet-3D/}.

Yang Chen, Yingwei Pan, Yehao Li et al.

Recent remarkable advances in large-scale text-to-image diffusion models have inspired a significant breakthrough in text-to-3D generation, pursuing 3D content creation solely from a given text prompt. However, existing text-to-3D techniques lack a crucial ability in the creative process: interactively control and shape the synthetic 3D contents according to users' desired specifications (e.g., sketch). To alleviate this issue, we present the first attempt for text-to-3D generation conditioning on the additional hand-drawn sketch, namely Control3D, which enhances controllability for users. In particular, a 2D conditioned diffusion model (ControlNet) is remoulded to guide the learning of 3D scene parameterized as NeRF, encouraging each view of 3D scene aligned with the given text prompt and hand-drawn sketch. Moreover, we exploit a pre-trained differentiable photo-to-sketch model to directly estimate the sketch of the rendered image over synthetic 3D scene. Such estimated sketch along with each sampled view is further enforced to be geometrically consistent with the given sketch, pursuing better controllable text-to-3D generation. Through extensive experiments, we demonstrate that our proposal can generate accurate and faithful 3D scenes that align closely with the input text prompts and sketches.

Xin Jiang, Jingwen Chen, Yehao Li et al.

Recent advances in subject-driven image generation using diffusion models have attracted considerable attention for their remarkable capabilities in producing high-quality images. Nevertheless, the potential of Visual Autoregressive (VAR) models, despite their unified architecture and efficient inference, remains underexplored. In this work, we present DreamVAR, a novel framework for subject-driven image synthesis built upon a VAR model that employs next-scale prediction. Technically, multi-scale features of the reference subject are first extracted by a visual tokenizer. Instead of interleaving these conditional features with target image tokens across scales, our DreamVAR pre-fills the full subject feature sequence prior to predicting target image tokens. This design simplifies autoregressive dependencies and mitigates the train-test discrepancy in multi-scale conditioning scenario within the VAR paradigm. DreamVAR further incorporates reinforcement learning to jointly enhance semantic alignment and subject consistency. Extensive experiments demonstrate that DreamVAR achieves superior appearance preservation compared to leading diffusion-based methods.

Qi Cai, Jingwen Chen, Yang Chen et al.

Recent advancements in image generative foundation models have prioritized quality improvements but often at the cost of increased computational complexity and inference latency. To address this critical trade-off, we introduce HiDream-I1, a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds. HiDream-I1 is constructed with a new sparse Diffusion Transformer (DiT) structure. Specifically, it starts with a dual-stream decoupled design of sparse DiT with dynamic Mixture-of-Experts (MoE) architecture, in which two separate encoders are first involved to independently process image and text tokens. Then, a single-stream sparse DiT structure with dynamic MoE architecture is adopted to trigger multi-model interaction for image generation in a cost-efficient manner. To support flexiable accessibility with varied model capabilities, we provide HiDream-I1 in three variants: HiDream-I1-Full, HiDream-I1-Dev, and HiDream-I1-Fast. Furthermore, we go beyond the typical text-to-image generation and remould HiDream-I1 with additional image conditions to perform precise, instruction-based editing on given images, yielding a new instruction-based image editing model namely HiDream-E1. Ultimately, by integrating text-to-image generation and instruction-based image editing, HiDream-I1 evolves to form a comprehensive image agent (HiDream-A1) capable of fully interactive image creation and refinement. To accelerate multi-modal AIGC research, we have open-sourced all the codes and model weights of HiDream-I1-Full, HiDream-I1-Dev, HiDream-I1-Fast, HiDream-E1 through our project websites: https://github.com/HiDream-ai/HiDream-I1 and https://github.com/HiDream-ai/HiDream-E1. All features can be directly experienced via https://vivago.ai/studio.

Qingyang Mao, Qi Cai, Yehao Li et al.

Recent advances in diffusion models have brought remarkable visual fidelity to instruction-guided image editing. However, their global denoising process inherently entangles the edited region with the entire image context, leading to unintended spurious modifications and compromised adherence to editing instructions. In contrast, autoregressive models offer a distinct paradigm by formulating image synthesis as a sequential process over discrete visual tokens. Their causal and compositional mechanism naturally circumvents the adherence challenges of diffusion-based methods. In this paper, we present VAREdit, a visual autoregressive (VAR) framework that reframes image editing as a next-scale prediction problem. Conditioned on source image features and text instructions, VAREdit generates multi-scale target features to achieve precise edits. A core challenge in this paradigm is how to effectively condition the source image tokens. We observe that finest-scale source features cannot effectively guide the prediction of coarser target features. To bridge this gap, we introduce a Scale-Aligned Reference (SAR) module, which injects scale-matched conditioning information into the first self-attention layer. VAREdit demonstrates significant advancements in both editing adherence and efficiency. On standard benchmarks, it outperforms leading diffusion-based methods by 30\%+ higher GPT-Balance score. Moreover, it completes a $512\times512$ editing in 1.2 seconds, making it 2.2$\times$ faster than the similarly sized UltraEdit. The models are available at https://github.com/HiDream-ai/VAREdit.

Qi Cai, Jingwen Chen, Chengmin Gao et al.

The evolution of visual generative models has long been constrained by fragmented architectures relying on disjoint text encoders and external VAEs. In this report, we present HiDream-O1-Image, a natively unified generative foundation model via pixel-space Diffusion Transformer, that pioneers a paradigm shift from modular architectures to an end-to-end in-context visual generation engine. By mapping raw image pixels, text tokens, and task-specific conditions into a single shared token space, HiDream-O1-Image achieves a structural unification of multimodal inputs within an Unified Transformer (UiT) architecture. This native encoding paradigm eliminates the need for separate VAEs or disjoint pre-trained text encoders, allowing the model to treat diverse generation and editing tasks as a consistent in-context reasoning process. Extensive experiments show that HiDream-O1-Image excels across various generation tasks, including text-to-image generation, instruction-based editing, and subject-driven personalization. Notably, with only 8B parameters, HiDream-O1-Image (8B) achieves performance parity with or even surpasses established state-of-the-art models with significantly larger parameters (e.g., 27B Qwen-Image). Crucially, to validate the immense scalability of this paradigm, we successfully scale the architecture up to over 200B parameters. Experimental results demonstrate that this massive-scale version HiDream-O1-Image-Pro (200B+) unlocks unprecedented generative capabilities and superior performance, establishing new state-of-the-art benchmarks. Ultimately, HiDream-O1-Image highlights the immense potential of natively unified architectures and charts a highly scalable path toward next-generation multimodal AI.

Yehao Li, Yingwei Pan, Jingwen Chen et al.

With the rise and development of deep learning over the past decade, there has been a steady momentum of innovation and breakthroughs that convincingly push the state-of-the-art of cross-modal analytics between vision and language in multimedia field. Nevertheless, there has not been an open-source codebase in support of training and deploying numerous neural network models for cross-modal analytics in a unified and modular fashion. In this work, we propose X-modaler -- a versatile and high-performance codebase that encapsulates the state-of-the-art cross-modal analytics into several general-purpose stages (e.g., pre-processing, encoder, cross-modal interaction, decoder, and decode strategy). Each stage is empowered with the functionality that covers a series of modules widely adopted in state-of-the-arts and allows seamless switching in between. This way naturally enables a flexible implementation of state-of-the-art algorithms for image captioning, video captioning, and vision-language pre-training, aiming to facilitate the rapid development of research community. Meanwhile, since the effective modular designs in several stages (e.g., cross-modal interaction) are shared across different vision-language tasks, X-modaler can be simply extended to power startup prototypes for other tasks in cross-modal analytics, including visual question answering, visual commonsense reasoning, and cross-modal retrieval. X-modaler is an Apache-licensed codebase, and its source codes, sample projects and pre-trained models are available on-line: https://github.com/YehLi/xmodaler.

Yehao Li, Ting Yao, Yingwei Pan et al.

Transformer with self-attention has led to the revolutionizing of natural language processing field, and recently inspires the emergence of Transformer-style architecture design with competitive results in numerous computer vision tasks. Nevertheless, most of existing designs directly employ self-attention over a 2D feature map to obtain the attention matrix based on pairs of isolated queries and keys at each spatial location, but leave the rich contexts among neighbor keys under-exploited. In this work, we design a novel Transformer-style module, i.e., Contextual Transformer (CoT) block, for visual recognition. Such design fully capitalizes on the contextual information among input keys to guide the learning of dynamic attention matrix and thus strengthens the capacity of visual representation. Technically, CoT block first contextually encodes input keys via a $3\times3$ convolution, leading to a static contextual representation of inputs. We further concatenate the encoded keys with input queries to learn the dynamic multi-head attention matrix through two consecutive $1\times1$ convolutions. The learnt attention matrix is multiplied by input values to achieve the dynamic contextual representation of inputs. The fusion of the static and dynamic contextual representations are finally taken as outputs. Our CoT block is appealing in the view that it can readily replace each $3\times3$ convolution in ResNet architectures, yielding a Transformer-style backbone named as Contextual Transformer Networks (CoTNet). Through extensive experiments over a wide range of applications (e.g., image recognition, object detection and instance segmentation), we validate the superiority of CoTNet as a stronger backbone. Source code is available at \url{https://github.com/JDAI-CV/CoTNet}.

Yehao Li, Yingwei Pan, Ting Yao et al.

Despite having impressive vision-language (VL) pretraining with BERT-based encoder for VL understanding, the pretraining of a universal encoder-decoder for both VL understanding and generation remains challenging. The difficulty originates from the inherently different peculiarities of the two disciplines, e.g., VL understanding tasks capitalize on the unrestricted message passing across modalities, while generation tasks only employ visual-to-textual message passing. In this paper, we start with a two-stream decoupled design of encoder-decoder structure, in which two decoupled cross-modal encoder and decoder are involved to separately perform each type of proxy tasks, for simultaneous VL understanding and generation pretraining. Moreover, for VL pretraining, the dominant way is to replace some input visual/word tokens with mask tokens and enforce the multi-modal encoder/decoder to reconstruct the original tokens, but no mask token is involved when fine-tuning on downstream tasks. As an alternative, we propose a primary scheduled sampling strategy that elegantly mitigates such discrepancy via pretraining encoder-decoder in a two-pass manner. Extensive experiments demonstrate the compelling generalizability of our pretrained encoder-decoder by fine-tuning on four VL understanding and generation downstream tasks. Source code is available at \url{https://github.com/YehLi/TDEN}.

Yingwei Pan, Ting Yao, Yehao Li et al.

Recent progress on fine-grained visual recognition and visual question answering has featured Bilinear Pooling, which effectively models the 2$^{nd}$ order interactions across multi-modal inputs. Nevertheless, there has not been evidence in support of building such interactions concurrently with attention mechanism for image captioning. In this paper, we introduce a unified attention block -- X-Linear attention block, that fully employs bilinear pooling to selectively capitalize on visual information or perform multi-modal reasoning. Technically, X-Linear attention block simultaneously exploits both the spatial and channel-wise bilinear attention distributions to capture the 2$^{nd}$ order interactions between the input single-modal or multi-modal features. Higher and even infinity order feature interactions are readily modeled through stacking multiple X-Linear attention blocks and equipping the block with Exponential Linear Unit (ELU) in a parameter-free fashion, respectively. Furthermore, we present X-Linear Attention Networks (dubbed as X-LAN) that novelly integrates X-Linear attention block(s) into image encoder and sentence decoder of image captioning model to leverage higher order intra- and inter-modal interactions. The experiments on COCO benchmark demonstrate that our X-LAN obtains to-date the best published CIDEr performance of 132.0% on COCO Karpathy test split. When further endowing Transformer with X-Linear attention blocks, CIDEr is boosted up to 132.8%. Source code is available at \url{https://github.com/Panda-Peter/image-captioning}.

Yingwei Pan, Yehao Li, Qi Cai et al.

This notebook paper presents an overview and comparative analysis of our systems designed for the following two tasks in Visual Domain Adaptation Challenge (VisDA-2019): multi-source domain adaptation and semi-supervised domain adaptation. Multi-Source Domain Adaptation: We investigate both pixel-level and feature-level adaptation for multi-source domain adaptation task, i.e., directly hallucinating labeled target sample via CycleGAN and learning domain-invariant feature representations through self-learning. Moreover, the mechanism of fusing features from different backbones is further studied to facilitate the learning of domain-invariant classifiers. Source code and pre-trained models are available at \url{https://github.com/Panda-Peter/visda2019-multisource}. Semi-Supervised Domain Adaptation: For this task, we adopt a standard self-learning framework to construct a classifier based on the labeled source and target data, and generate the pseudo labels for unlabeled target data. These target data with pseudo labels are then exploited to re-training the classifier in a following iteration. Furthermore, a prototype-based classification module is additionally utilized to strengthen the predictions. Source code and pre-trained models are available at \url{https://github.com/Panda-Peter/visda2019-semisupervised}.

Yurui Qian, Qi Cai, Yingwei Pan et al.

Diffusion models have recently brought a powerful revolution in image generation. Despite showing impressive generative capabilities, most of these models rely on the current sample to denoise the next one, possibly resulting in denoising instability. In this paper, we reinterpret the iterative denoising process as model optimization and leverage a moving average mechanism to ensemble all the prior samples. Instead of simply applying moving average to the denoised samples at different timesteps, we first map the denoised samples to data space and then perform moving average to avoid distribution shift across timesteps. In view that diffusion models evolve the recovery from low-frequency components to high-frequency details, we further decompose the samples into different frequency components and execute moving average separately on each component. We name the complete approach "Moving Average Sampling in Frequency domain (MASF)". MASF could be seamlessly integrated into mainstream pre-trained diffusion models and sampling schedules. Extensive experiments on both unconditional and conditional diffusion models demonstrate that our MASF leads to superior performances compared to the baselines, with almost negligible additional complexity cost.

Ting Yao, Yehao Li, Yingwei Pan et al.

The hybrid deep models of Vision Transformer (ViT) and Convolution Neural Network (CNN) have emerged as a powerful class of backbones for vision tasks. Scaling up the input resolution of such hybrid backbones naturally strengthes model capacity, but inevitably suffers from heavy computational cost that scales quadratically. Instead, we present a new hybrid backbone with HIgh-Resolution Inputs (namely HIRI-ViT), that upgrades prevalent four-stage ViT to five-stage ViT tailored for high-resolution inputs. HIRI-ViT is built upon the seminal idea of decomposing the typical CNN operations into two parallel CNN branches in a cost-efficient manner. One high-resolution branch directly takes primary high-resolution features as inputs, but uses less convolution operations. The other low-resolution branch first performs down-sampling and then utilizes more convolution operations over such low-resolution features. Experiments on both recognition task (ImageNet-1K dataset) and dense prediction tasks (COCO and ADE20K datasets) demonstrate the superiority of HIRI-ViT. More remarkably, under comparable computational cost ($\sim$5.0 GFLOPs), HIRI-ViT achieves to-date the best published Top-1 accuracy of 84.3% on ImageNet with 448$\times$448 inputs, which absolutely improves 83.4% of iFormer-S by 0.9% with 224$\times$224 inputs.

Rui Zhu, Yingwei Pan, Yehao Li et al.

Diffusion Transformer (DiT) has emerged as the new trend of generative diffusion models on image generation. In view of extremely slow convergence in typical DiT, recent breakthroughs have been driven by mask strategy that significantly improves the training efficiency of DiT with additional intra-image contextual learning. Despite this progress, mask strategy still suffers from two inherent limitations: (a) training-inference discrepancy and (b) fuzzy relations between mask reconstruction & generative diffusion process, resulting in sub-optimal training of DiT. In this work, we address these limitations by novelly unleashing the self-supervised discrimination knowledge to boost DiT training. Technically, we frame our DiT in a teacher-student manner. The teacher-student discriminative pairs are built on the diffusion noises along the same Probability Flow Ordinary Differential Equation (PF-ODE). Instead of applying mask reconstruction loss over both DiT encoder and decoder, we decouple DiT encoder and decoder to separately tackle discriminative and generative objectives. In particular, by encoding discriminative pairs with student and teacher DiT encoders, a new discriminative loss is designed to encourage the inter-image alignment in the self-supervised embedding space. After that, student samples are fed into student DiT decoder to perform the typical generative diffusion task. Extensive experiments are conducted on ImageNet dataset, and our method achieves a competitive balance between training cost and generative capacity.

Jianjie Luo, Jingwen Chen, Yehao Li et al.

Recently, zero-shot image captioning has gained increasing attention, where only text data is available for training. The remarkable progress in text-to-image diffusion model presents the potential to resolve this task by employing synthetic image-caption pairs generated by this pre-trained prior. Nonetheless, the defective details in the salient regions of the synthetic images introduce semantic misalignment between the synthetic image and text, leading to compromised results. To address this challenge, we propose a novel Patch-wise Cross-modal feature Mix-up (PCM) mechanism to adaptively mitigate the unfaithful contents in a fine-grained manner during training, which can be integrated into most of encoder-decoder frameworks, introducing our PCM-Net. Specifically, for each input image, salient visual concepts in the image are first detected considering the image-text similarity in CLIP space. Next, the patch-wise visual features of the input image are selectively fused with the textual features of the salient visual concepts, leading to a mixed-up feature map with less defective content. Finally, a visual-semantic encoder is exploited to refine the derived feature map, which is further incorporated into the sentence decoder for caption generation. Additionally, to facilitate the model training with synthetic data, a novel CLIP-weighted cross-entropy loss is devised to prioritize the high-quality image-text pairs over the low-quality counterparts. Extensive experiments on MSCOCO and Flickr30k datasets demonstrate the superiority of our PCM-Net compared with state-of-the-art VLMs-based approaches. It is noteworthy that our PCM-Net ranks first in both in-domain and cross-domain zero-shot image captioning. The synthetic dataset SynthImgCap and code are available at https://jianjieluo.github.io/SynthImgCap.

Yehao Li, Jiahao Fan, Yingwei Pan et al.

Vision-language pre-training has been an emerging and fast-developing research topic, which transfers multi-modal knowledge from rich-resource pre-training task to limited-resource downstream tasks. Unlike existing works that predominantly learn a single generic encoder, we present a pre-trainable Universal Encoder-DEcoder Network (Uni-EDEN) to facilitate both vision-language perception (e.g., visual question answering) and generation (e.g., image captioning). Uni-EDEN is a two-stream Transformer based structure, consisting of three modules: object and sentence encoders that separately learns the representations of each modality, and sentence decoder that enables both multi-modal reasoning and sentence generation via inter-modal interaction. Considering that the linguistic representations of each image can span different granularities in this hierarchy including, from simple to comprehensive, individual label, a phrase, and a natural sentence, we pre-train Uni-EDEN through multi-granular vision-language proxy tasks: Masked Object Classification (MOC), Masked Region Phrase Generation (MRPG), Image-Sentence Matching (ISM), and Masked Sentence Generation (MSG). In this way, Uni-EDEN is endowed with the power of both multi-modal representation extraction and language modeling. Extensive experiments demonstrate the compelling generalizability of Uni-EDEN by fine-tuning it to four vision-language perception and generation downstream tasks.

Jianjie Luo, Yehao Li, Yingwei Pan et al.

BERT-type structure has led to the revolution of vision-language pre-training and the achievement of state-of-the-art results on numerous vision-language downstream tasks. Existing solutions dominantly capitalize on the multi-modal inputs with mask tokens to trigger mask-based proxy pre-training tasks (e.g., masked language modeling and masked object/frame prediction). In this work, we argue that such masked inputs would inevitably introduce noise for cross-modal matching proxy task, and thus leave the inherent vision-language association under-explored. As an alternative, we derive a particular form of cross-modal proxy objective for video-language pre-training, i.e., Contrastive Cross-modal matching and denoising (CoCo). By viewing the masked frame/word sequences as the noisy augmentation of primary unmasked ones, CoCo strengthens video-language association by simultaneously pursuing inter-modal matching and intra-modal denoising between masked and unmasked inputs in a contrastive manner. Our CoCo proxy objective can be further integrated into any BERT-type encoder-decoder structure for video-language pre-training, named as Contrastive Cross-modal BERT (CoCo-BERT). We pre-train CoCo-BERT on TV dataset and a newly collected large-scale GIF video dataset (ACTION). Through extensive experiments over a wide range of downstream tasks (e.g., cross-modal retrieval, video question answering, and video captioning), we demonstrate the superiority of CoCo-BERT as a pre-trained structure.

Yingwei Pan, Jun Xu, Yehao Li et al.

The Pre-training for Video Captioning Challenge 2020 Summary: results and challenge participants' technical reports.

Yingwei Pan, Yehao Li, Jianjie Luo et al.

In this work, we present Auto-captions on GIF, which is a new large-scale pre-training dataset for generic video understanding. All video-sentence pairs are created by automatically extracting and filtering video caption annotations from billions of web pages. Auto-captions on GIF dataset can be utilized to pre-train the generic feature representation or encoder-decoder structure for video captioning, and other downstream tasks (e.g., sentence localization in videos, video question answering, etc.) as well. We present a detailed analysis of Auto-captions on GIF dataset in comparison to existing video-sentence datasets. We also provide an evaluation of a Transformer-based encoder-decoder structure for vision-language pre-training, which is further adapted to video captioning downstream task and yields the compelling generalizability on MSR-VTT. The dataset is available at \url{http://www.auto-video-captions.top/2020/dataset}.

Yingwei Pan, Ting Yao, Yehao Li et al.

Unsupervised domain adaptation has received significant attention in recent years. Most of existing works tackle the closed-set scenario, assuming that the source and target domains share the exactly same categories. In practice, nevertheless, a target domain often contains samples of classes unseen in source domain (i.e., unknown class). The extension of domain adaptation from closed-set to such open-set situation is not trivial since the target samples in unknown class are not expected to align with the source. In this paper, we address this problem by augmenting the state-of-the-art domain adaptation technique, Self-Ensembling, with category-agnostic clusters in target domain. Specifically, we present Self-Ensembling with Category-agnostic Clusters (SE-CC) -- a novel architecture that steers domain adaptation with the additional guidance of category-agnostic clusters that are specific to target domain. These clustering information provides domain-specific visual cues, facilitating the generalization of Self-Ensembling for both closed-set and open-set scenarios. Technically, clustering is firstly performed over all the unlabeled target samples to obtain the category-agnostic clusters, which reveal the underlying data space structure peculiar to target domain. A clustering branch is capitalized on to ensure that the learnt representation preserves such underlying structure by matching the estimated assignment distribution over clusters to the inherent cluster distribution for each target sample. Furthermore, SE-CC enhances the learnt representation with mutual information maximization. Extensive experiments are conducted on Office and VisDA datasets for both open-set and closed-set domain adaptation, and superior results are reported when comparing to the state-of-the-art approaches.

Ting Yao, Yingwei Pan, Yehao Li et al.

It is always well believed that parsing an image into constituent visual patterns would be helpful for understanding and representing an image. Nevertheless, there has not been evidence in support of the idea on describing an image with a natural-language utterance. In this paper, we introduce a new design to model a hierarchy from instance level (segmentation), region level (detection) to the whole image to delve into a thorough image understanding for captioning. Specifically, we present a HIerarchy Parsing (HIP) architecture that novelly integrates hierarchical structure into image encoder. Technically, an image decomposes into a set of regions and some of the regions are resolved into finer ones. Each region then regresses to an instance, i.e., foreground of the region. Such process naturally builds a hierarchal tree. A tree-structured Long Short-Term Memory (Tree-LSTM) network is then employed to interpret the hierarchal structure and enhance all the instance-level, region-level and image-level features. Our HIP is appealing in view that it is pluggable to any neural captioning models. Extensive experiments on COCO image captioning dataset demonstrate the superiority of HIP. More remarkably, HIP plus a top-down attention-based LSTM decoder increases CIDEr-D performance from 120.1% to 127.2% on COCO Karpathy test split. When further endowing instance-level and region-level features from HIP with semantic relation learnt through Graph Convolutional Networks (GCN), CIDEr-D is boosted up to 130.6%.

Yehao Li, Ting Yao, Yingwei Pan et al.

The problem of distance metric learning is mostly considered from the perspective of learning an embedding space, where the distances between pairs of examples are in correspondence with a similarity metric. With the rise and success of Convolutional Neural Networks (CNN), deep metric learning (DML) involves training a network to learn a nonlinear transformation to the embedding space. Existing DML approaches often express the supervision through maximizing inter-class distance and minimizing intra-class variation. However, the results can suffer from overfitting problem, especially when the training examples of each class are embedded together tightly and the density of each class is very high. In this paper, we integrate density, i.e., the measure of data concentration in the representation, into the optimization of DML frameworks to adaptively balance inter-class similarity and intra-class variation by training the architecture in an end-to-end manner. Technically, the knowledge of density is employed as a regularizer, which is pluggable to any DML architecture with different objective functions such as contrastive loss, N-pair loss and triplet loss. Extensive experiments on three public datasets consistently demonstrate clear improvements by amending three types of embedding with the density adaptivity. More remarkably, our proposal increases Recall@1 from 67.95% to 77.62%, from 52.01% to 55.64% and from 68.20% to 70.56% on Cars196, CUB-200-2011 and Stanford Online Products dataset, respectively.

Zhaofan Qiu, Dong Li, Yehao Li et al.

This notebook paper presents an overview and comparative analysis of our systems designed for the following three tasks in ActivityNet Challenge 2019: trimmed action recognition, dense-captioning events in videos, and spatio-temporal action localization.

Jingwen Chen, Yingwei Pan, Yehao Li et al.

It is well believed that video captioning is a fundamental but challenging task in both computer vision and artificial intelligence fields. The prevalent approach is to map an input video to a variable-length output sentence in a sequence to sequence manner via Recurrent Neural Network (RNN). Nevertheless, the training of RNN still suffers to some degree from vanishing/exploding gradient problem, making the optimization difficult. Moreover, the inherently recurrent dependency in RNN prevents parallelization within a sequence during training and therefore limits the computations. In this paper, we present a novel design --- Temporal Deformable Convolutional Encoder-Decoder Networks (dubbed as TDConvED) that fully employ convolutions in both encoder and decoder networks for video captioning. Technically, we exploit convolutional block structures that compute intermediate states of a fixed number of inputs and stack several blocks to capture long-term relationships. The structure in encoder is further equipped with temporal deformable convolution to enable free-form deformation of temporal sampling. Our model also capitalizes on temporal attention mechanism for sentence generation. Extensive experiments are conducted on both MSVD and MSR-VTT video captioning datasets, and superior results are reported when comparing to conventional RNN-based encoder-decoder techniques. More remarkably, TDConvED increases CIDEr-D performance from 58.8% to 67.2% on MSVD.

Yehao Li, Ting Yao, Yingwei Pan et al.

Image captioning has received significant attention with remarkable improvements in recent advances. Nevertheless, images in the wild encapsulate rich knowledge and cannot be sufficiently described with models built on image-caption pairs containing only in-domain objects. In this paper, we propose to address the problem by augmenting standard deep captioning architectures with object learners. Specifically, we present Long Short-Term Memory with Pointing (LSTM-P) --- a new architecture that facilitates vocabulary expansion and produces novel objects via pointing mechanism. Technically, object learners are initially pre-trained on available object recognition data. Pointing in LSTM-P then balances the probability between generating a word through LSTM and copying a word from the recognized objects at each time step in decoder stage. Furthermore, our captioning encourages global coverage of objects in the sentence. Extensive experiments are conducted on both held-out COCO image captioning and ImageNet datasets for describing novel objects, and superior results are reported when comparing to state-of-the-art approaches. More remarkably, we obtain an average of 60.9% in F1 score on held-out COCO~dataset.

Yingwei Pan, Ting Yao, Yehao Li et al.

In this paper, we introduce a new idea for unsupervised domain adaptation via a remold of Prototypical Networks, which learn an embedding space and perform classification via a remold of the distances to the prototype of each class. Specifically, we present Transferrable Prototypical Networks (TPN) for adaptation such that the prototypes for each class in source and target domains are close in the embedding space and the score distributions predicted by prototypes separately on source and target data are similar. Technically, TPN initially matches each target example to the nearest prototype in the source domain and assigns an example a "pseudo" label. The prototype of each class could then be computed on source-only, target-only and source-target data, respectively. The optimization of TPN is end-to-end trained by jointly minimizing the distance across the prototypes on three types of data and KL-divergence of score distributions output by each pair of the prototypes. Extensive experiments are conducted on the transfers across MNIST, USPS and SVHN datasets, and superior results are reported when comparing to state-of-the-art approaches. More remarkably, we obtain an accuracy of 80.4% of single model on VisDA 2017 dataset.

Ting Yao, Yingwei Pan, Yehao Li et al.

It is always well believed that modeling relationships between objects would be helpful for representing and eventually describing an image. Nevertheless, there has not been evidence in support of the idea on image description generation. In this paper, we introduce a new design to explore the connections between objects for image captioning under the umbrella of attention-based encoder-decoder framework. Specifically, we present Graph Convolutional Networks plus Long Short-Term Memory (dubbed as GCN-LSTM) architecture that novelly integrates both semantic and spatial object relationships into image encoder. Technically, we build graphs over the detected objects in an image based on their spatial and semantic connections. The representations of each region proposed on objects are then refined by leveraging graph structure through GCN. With the learnt region-level features, our GCN-LSTM capitalizes on LSTM-based captioning framework with attention mechanism for sentence generation. Extensive experiments are conducted on COCO image captioning dataset, and superior results are reported when comparing to state-of-the-art approaches. More remarkably, GCN-LSTM increases CIDEr-D performance from 120.1% to 128.7% on COCO testing set.

Yehao Li, Ting Yao, Yingwei Pan et al.

Automatically describing a video with natural language is regarded as a fundamental challenge in computer vision. The problem nevertheless is not trivial especially when a video contains multiple events to be worthy of mention, which often happens in real videos. A valid question is how to temporally localize and then describe events, which is known as "dense video captioning." In this paper, we present a novel framework for dense video captioning that unifies the localization of temporal event proposals and sentence generation of each proposal, by jointly training them in an end-to-end manner. To combine these two worlds, we integrate a new design, namely descriptiveness regression, into a single shot detection structure to infer the descriptive complexity of each detected proposal via sentence generation. This in turn adjusts the temporal locations of each event proposal. Our model differs from existing dense video captioning methods since we propose a joint and global optimization of detection and captioning, and the framework uniquely capitalizes on an attribute-augmented video captioning architecture. Extensive experiments are conducted on ActivityNet Captions dataset and our framework shows clear improvements when compared to the state-of-the-art techniques. More remarkably, we obtain a new record: METEOR of 12.96% on ActivityNet Captions official test set.

Ting Yao, Yingwei Pan, Yehao Li et al.

Image captioning often requires a large set of training image-sentence pairs. In practice, however, acquiring sufficient training pairs is always expensive, making the recent captioning models limited in their ability to describe objects outside of training corpora (i.e., novel objects). In this paper, we present Long Short-Term Memory with Copying Mechanism (LSTM-C) --- a new architecture that incorporates copying into the Convolutional Neural Networks (CNN) plus Recurrent Neural Networks (RNN) image captioning framework, for describing novel objects in captions. Specifically, freely available object recognition datasets are leveraged to develop classifiers for novel objects. Our LSTM-C then nicely integrates the standard word-by-word sentence generation by a decoder RNN with copying mechanism which may instead select words from novel objects at proper places in the output sentence. Extensive experiments are conducted on both MSCOCO image captioning and ImageNet datasets, demonstrating the ability of our proposed LSTM-C architecture to describe novel objects. Furthermore, superior results are reported when compared to state-of-the-art deep models.

Ting Yao, Yingwei Pan, Yehao Li et al.

Automatically describing an image with a natural language has been an emerging challenge in both fields of computer vision and natural language processing. In this paper, we present Long Short-Term Memory with Attributes (LSTM-A) - a novel architecture that integrates attributes into the successful Convolutional Neural Networks (CNNs) plus Recurrent Neural Networks (RNNs) image captioning framework, by training them in an end-to-end manner. To incorporate attributes, we construct variants of architectures by feeding image representations and attributes into RNNs in different ways to explore the mutual but also fuzzy relationship between them. Extensive experiments are conducted on COCO image captioning dataset and our framework achieves superior results when compared to state-of-the-art deep models. Most remarkably, we obtain METEOR/CIDEr-D of 25.2%/98.6% on testing data of widely used and publicly available splits in (Karpathy & Fei-Fei, 2015) when extracting image representations by GoogleNet and achieve to date top-1 performance on COCO captioning Leaderboard.