Video summarization mainly aims to produce a compact, short, informative, and representative synopsis of raw videos, which is of great importance for browsing, analyzing, and understanding video content. Dominant video summarization approaches are generally based on recurrent or convolutional neural networks, even recent encoder-only transformers. We propose using full transformer as an alternative architecture to perform video summarization. The full transformer with an encoder-decoder structure, specifically designed for handling sequence transduction problems, is naturally suitable for video summarization tasks. This work considers supervised video summarization and casts it as a sequence-to-sequence learning problem. Our key idea is to directly apply the full transformer to the video summarization task, which is intuitively sound and effective. Also, considering the efficiency problem, we replace full attention with the combination of local and global sparse attention, which enables modeling long-range dependencies while reducing computational costs. Based on this, we propose a transformer-like architecture, named FullTransNet, which has a full encoder-decoder structure with local-global sparse attention for video summarization. Specifically, both the encoder and decoder in FullTransNet are stacked the same way as ones in the vanilla transformer, and the local-global sparse attention is used only at the encoder side. Extensive experiments on two public multimedia benchmark datasets SumMe and TVSum demonstrate that our proposed model can outperform other video summarization approaches, achieving F-Measures of 54.4% on SumMe and 63.9% on TVSum with relatively lower compute and memory requirements, verifying its effectiveness and efficiency. The code and models are publicly available on GitHub.
Exploring the bridge between historical and future motion behaviors remains a central challenge in human motion prediction. While most existing methods incorporate a reconstruction task as an auxiliary task into the decoder, thereby improving the modeling of spatio-temporal dependencies, they overlook the potential conflicts between reconstruction and prediction tasks. In this paper, we propose a novel approach: Temporal Decoupling Decoding with Inverse Processing (\textbf{$TD^2IP$}). Our method strategically separates reconstruction and prediction decoding processes, employing distinct decoders to decode the shared motion features into historical or future sequences. Additionally, inverse processing reverses motion information in the temporal dimension and reintroduces it into the model, leveraging the bidirectional temporal correlation of human motion behaviors. By alleviating the conflicts between reconstruction and prediction tasks and enhancing the association of historical and future information, \textbf{$TD^2IP$} fosters a deeper understanding of motion patterns. Extensive experiments demonstrate the adaptability of our method within existing methods.
Diffusion models have shown exceptional performance in visual generation tasks. Recently, these models have shifted from traditional U-Shaped CNN-Attention hybrid structures to fully transformer-based isotropic architectures. While these transformers exhibit strong scalability and performance, their reliance on complicated self-attention operation results in slow inference speeds. Contrary to these works, we rethink one of the simplest yet fastest module in deep learning, 3x3 Convolution, to construct a scaled-up purely convolutional diffusion model. We first discover that an Encoder-Decoder Hourglass design outperforms scalable isotropic architectures for Conv3x3, but still under-performing our expectation. Further improving the architecture, we introduce sparse skip connections to reduce redundancy and improve scalability. Based on the architecture, we introduce conditioning improvements including stage-specific embeddings, mid-block condition injection, and conditional gating. These improvements lead to our proposed Diffusion CNN (DiC), which serves as a swift yet competitive diffusion architecture baseline. Experiments on various scales and settings show that DiC surpasses existing diffusion transformers by considerable margins in terms of performance while keeping a good speed advantage. Project page: https://github.com/YuchuanTian/DiC
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.
Existing unsupervised distillation-based methods rely on the differences between encoded and decoded features to locate abnormal regions in test images. However, the decoder trained only on normal samples still reconstructs abnormal patch features well, degrading performance. This issue is particularly pronounced in unsupervised multi-class anomaly detection tasks. We attribute this behavior to over-generalization(OG) of decoder: the significantly increasing diversity of patch patterns in multi-class training enhances the model generalization on normal patches, but also inadvertently broadens its generalization to abnormal patches. To mitigate OG, we propose a novel approach that leverages class-agnostic learnable prompts to capture common textual normality across various visual patterns, and then apply them to guide the decoded features towards a normal textual representation, suppressing over-generalization of the decoder on abnormal patterns. To further improve performance, we also introduce a gated mixture-of-experts module to specialize in handling diverse patch patterns and reduce mutual interference between them in multi-class training. Our method achieves competitive performance on the MVTec AD and VisA datasets, demonstrating its effectiveness.
Text-based person search is the task of finding person images that are the most relevant to the natural language text description given as query. The main challenge of this task is a large gap between the target images and text queries, which makes it difficult to establish correspondence and distinguish subtle differences across people. To address this challenge, we introduce an efficient encoder-decoder model that extracts coarse-to-fine embedding vectors which are semantically aligned across the two modalities without supervision for the alignment. There is another challenge of learning to capture fine-grained information with only person IDs as supervision, where similar body parts of different individuals are considered different due to the lack of part-level supervision. To tackle this, we propose a novel ranking loss, dubbed commonality-based margin ranking loss, which quantifies the degree of commonality of each body part and reflects it during the learning of fine-grained body part details. As a consequence, it enables our method to achieve the best records on three public benchmarks.
Existing neural methods for the Travelling Salesman Problem (TSP) mostly aim at finding a single optimal solution. To discover diverse yet high-quality solutions for Multi-Solution TSP (MSTSP), we propose a novel deep reinforcement learning based neural solver, which is primarily featured by an encoder-decoder structured policy. Concretely, on the one hand, a Relativization Filter (RF) is designed to enhance the robustness of the encoder to affine transformations of the instances, so as to potentially improve the quality of the found solutions. On the other hand, a Multi-Attentive Adaptive Active Search (MA3S) is tailored to allow the decoders to strike a balance between the optimality and diversity. Experimental evaluations on benchmark instances demonstrate the superiority of our method over recent neural baselines across different metrics, and its competitive performance against state-of-the-art traditional heuristics with significantly reduced computational time, ranging from $1.3\times$ to $15\times$ faster. Furthermore, we demonstrate that our method can also be applied to the Capacitated Vehicle Routing Problem (CVRP).
The goal of voice conversion is to transform the speech of a source speaker to sound like that of a reference speaker while preserving the original content. A key challenge is to extract disentangled linguistic content from the source and voice style from the reference. While existing approaches leverage various methods to isolate the two, a generalization still requires further attention, especially for robustness in zero-shot scenarios. In this paper, we achieve successful disentanglement of content and speaker features by tuning self-supervised speech features with adapters. The adapters are trained to dynamically encode nuanced features from rich self-supervised features, and the decoder fuses them to produce speech that accurately resembles the reference with minimal loss of content. Moreover, we leverage a conditional flow matching decoder with cross-attention speaker conditioning to further boost the synthesis quality and efficiency. Subjective and objective evaluations in a zero-shot scenario demonstrate that the proposed method outperforms existing models in speech quality and similarity to the reference speech.
Current sketch extraction methods either require extensive training or fail to capture a wide range of artistic styles, limiting their practical applicability and versatility. We introduce Mixture-of-Self-Attention (MixSA), a training-free sketch extraction method that leverages strong diffusion priors for enhanced sketch perception. At its core, MixSA employs a mixture-of-self-attention technique, which manipulates self-attention layers by substituting the keys and values with those from reference sketches. This allows for the seamless integration of brushstroke elements into initial outline images, offering precise control over texture density and enabling interpolation between styles to create novel, unseen styles. By aligning brushstroke styles with the texture and contours of colored images, particularly in late decoder layers handling local textures, MixSA addresses the common issue of color averaging by adjusting initial outlines. Evaluated with various perceptual metrics, MixSA demonstrates superior performance in sketch quality, flexibility, and applicability. This approach not only overcomes the limitations of existing methods but also empowers users to generate diverse, high-fidelity sketches that more accurately reflect a wide range of artistic expressions.
Generative models have enabled easy creation and generation of images of all kinds given a single prompt. However, this has also raised ethical concerns about what is an actual piece of content created by humans or cameras compared to model-generated content like images or videos. Watermarking data generated by modern generative models is a popular method to provide information on the source of the content. The goal is for all generated images to conceal an invisible watermark, allowing for future detection or identification. The Stable Signature finetunes the decoder of Latent Diffusion Models such that a unique watermark is rooted in any image produced by the decoder. In this paper, we present a novel adversarial fine-tuning attack that disrupts the model's ability to embed the intended watermark, exposing a significant vulnerability in existing watermarking methods. To address this, we further propose a tamper-resistant fine-tuning algorithm inspired by methods developed for large language models, tailored to the specific requirements of watermarking in LDMs. Our findings emphasize the importance of anticipating and defending against potential vulnerabilities in generative systems.
Accurate 3D perception is essential for understanding the environment in autonomous driving. Recent advancements in 3D semantic occupancy prediction have leveraged camera-LiDAR fusion to improve robustness and accuracy. However, current methods allocate computational resources uniformly across all voxels, leading to inefficiency, and they also fail to adequately address occlusions, resulting in reduced accuracy in challenging scenarios. We propose MR-Occ, a novel approach for camera-LiDAR fusion-based 3D semantic occupancy prediction, addressing these challenges through three key components: Hierarchical Voxel Feature Refinement (HVFR), Multi-scale Occupancy Decoder (MOD), and Pixel to Voxel Fusion Network (PVF-Net). HVFR improves performance by enhancing features for critical voxels, reducing computational cost. MOD introduces an `occluded' class to better handle regions obscured from sensor view, improving accuracy. PVF-Net leverages densified LiDAR features to effectively fuse camera and LiDAR data through a deformable attention mechanism. Extensive experiments demonstrate that MR-Occ achieves state-of-the-art performance on the nuScenes-Occupancy dataset, surpassing previous approaches by +5.2% in IoU and +5.3% in mIoU while using fewer parameters and FLOPs. Moreover, MR-Occ demonstrates superior performance on the SemanticKITTI dataset, further validating its effectiveness and generalizability across diverse 3D semantic occupancy benchmarks.
Autonomous vehicles face significant challenges in navigating adverse weather, particularly rain, due to the visual impairment of camera-based systems. In this study, we leveraged contemporary deep learning techniques to mitigate these challenges, aiming to develop a vision model that processes live vehicle camera feeds to eliminate rain-induced visual hindrances, yielding visuals closely resembling clear, rain-free scenes. Using the Car Learning to Act (CARLA) simulation environment, we generated a comprehensive dataset of clear and rainy images for model training and testing. In our model, we employed a classic encoder-decoder architecture with skip connections and concatenation operations. It was trained using novel batching schemes designed to effectively distinguish high-frequency rain patterns from low-frequency scene features across successive image frames. To evaluate the model performance, we integrated it with a steering module that processes front-view images as input. The results demonstrated notable improvements in steering accuracy, underscoring the model's potential to enhance navigation safety and reliability in rainy weather conditions.
Within the evolving landscape of smart homes, the precise recognition of daily living activities using ambient sensor data stands paramount. This paper not only aims to bolster existing algorithms by evaluating two distinct pretrained embeddings suited for ambient sensor activations but also introduces a novel hierarchical architecture. We delve into an architecture anchored on Transformer Decoder-based pre-trained embeddings, reminiscent of the GPT design, and contrast it with the previously established state-of-the-art (SOTA) ELMo embeddings for ambient sensors. Our proposed hierarchical structure leverages the strengths of each pre-trained embedding, enabling the discernment of activity dependencies and sequence order, thereby enhancing classification precision. To further refine recognition, we incorporate into our proposed architecture an hour-of-the-day embedding. Empirical evaluations underscore the preeminence of the Transformer Decoder embedding in classification endeavors. Additionally, our innovative hierarchical design significantly bolsters the efficacy of both pre-trained embeddings, notably in capturing inter-activity nuances. The integration of temporal aspects subtly but distinctively augments classification, especially for time-sensitive activities. In conclusion, our GPT-inspired hierarchical approach, infused with temporal insights, outshines the SOTA ELMo benchmark.
This paper proposes a method for long-term action anticipation (LTA), the task of predicting action labels and their duration in a video given the observation of an initial untrimmed video interval. We build on an encoder-decoder architecture with parallel decoding and make two key contributions. First, we introduce a bi-directional action context regularizer module on the top of the decoder that ensures temporal context coherence in temporally adjacent segments. Second, we learn from classified segments a transition matrix that models the probability of transitioning from one action to another and the sequence is optimized globally over the full prediction interval. In addition, we use a specialized encoder for the task of action segmentation to increase the quality of the predictions in the observation interval at inference time, leading to a better understanding of the past. We validate our methods on four benchmark datasets for LTA, the EpicKitchen-55, EGTEA+, 50Salads and Breakfast demonstrating superior or comparable performance to state-of-the-art methods, including probabilistic models and also those based on Large Language Models, that assume trimmed video as input. The code will be released upon acceptance.
Informed by the success of the transformer model in various computer vision tasks, we design an end-to-end trainable model for the automatic detection and classification of bleeding and non-bleeding frames extracted from Wireless Capsule Endoscopy (WCE) videos. Based on the DETR model, our model uses the Resnet50 for feature extraction, the transformer encoder-decoder for bleeding and non-bleeding region detection, and a feedforward neural network for classification. Trained in an end-to-end approach on the Auto-WCEBleedGen Version 1 challenge training set, our model performs both detection and classification tasks as a single unit. Our model achieves an accuracy, recall, and F1-score classification percentage score of 98.28, 96.79, and 98.37 respectively, on the Auto-WCEBleedGen version 1 validation set. Further, we record an average precision (AP @ 0.5), mean-average precision (mAP) of 0.7447 and 0.7328 detection results. This earned us a 3rd place position in the challenge. Our code is publicly available via https://github.com/BasitAlawode/WCEBleedGen.
Visual Grounding aims to localize the referring object in an image given a natural language expression. Recent advancements in DETR-based visual grounding methods have attracted considerable attention, as they directly predict the coordinates of the target object without relying on additional efforts, such as pre-generated proposal candidates or pre-defined anchor boxes. However, existing research primarily focuses on designing stronger multi-modal decoder, which typically generates learnable queries by random initialization or by using linguistic embeddings. This vanilla query generation approach inevitably increases the learning difficulty for the model, as it does not involve any target-related information at the beginning of decoding. Furthermore, they only use the deepest image feature during the query learning process, overlooking the importance of features from other levels. To address these issues, we propose a novel approach, called RefFormer. It consists of the query adaption module that can be seamlessly integrated into CLIP and generate the referential query to provide the prior context for decoder, along with a task-specific decoder. By incorporating the referential query into the decoder, we can effectively mitigate the learning difficulty of the decoder, and accurately concentrate on the target object. Additionally, our proposed query adaption module can also act as an adapter, preserving the rich knowledge within CLIP without the need to tune the parameters of the backbone network. Extensive experiments demonstrate the effectiveness and efficiency of our proposed method, outperforming state-of-the-art approaches on five visual grounding benchmarks.
Point Transformers (PoinTr) have shown great potential in point cloud completion recently. Nevertheless, effective domain adaptation that improves transferability toward target domains remains unexplored. In this paper, we delve into this topic and empirically discover that direct feature alignment on point Transformer's CNN backbone only brings limited improvements since it cannot guarantee sequence-wise domain-invariant features in the Transformer. To this end, we propose a pioneering Domain Adaptive Point Transformer (DAPoinTr) framework for point cloud completion. DAPoinTr consists of three key components: Domain Query-based Feature Alignment (DQFA), Point Token-wise Feature alignment (PTFA), and Voted Prediction Consistency (VPC). In particular, DQFA is presented to narrow the global domain gaps from the sequence via the presented domain proxy and domain query at the Transformer encoder and decoder, respectively. PTFA is proposed to close the local domain shifts by aligning the tokens, \emph{i.e.,} point proxy and dynamic query, at the Transformer encoder and decoder, respectively. VPC is designed to consider different Transformer decoders as multiple of experts (MoE) for ensembled prediction voting and pseudo-label generation. Extensive experiments with visualization on several domain adaptation benchmarks demonstrate the effectiveness and superiority of our DAPoinTr compared with state-of-the-art methods. Code will be publicly available at: https://github.com/Yinghui-Li-New/DAPoinTr
As working with large datasets becomes standard, the task of accurately retrieving images containing objects of interest by an open set textual query gains practical importance. The current leading approach utilizes a pre-trained CLIP model without any adaptation to the target domain, balancing accuracy and efficiency through additional post-processing. In this work, we propose FOR: Finetuning for Object-centric Open-vocabulary Image Retrieval, which allows finetuning on a target dataset using closed-set labels while keeping the visual-language association crucial for open vocabulary retrieval. FOR is based on two design elements: a specialized decoder variant of the CLIP head customized for the intended task, and its coupling within a multi-objective training framework. Together, these design choices result in a significant increase in accuracy, showcasing improvements of up to 8 mAP@50 points over SoTA across three datasets. Additionally, we demonstrate that FOR is also effective in a semi-supervised setting, achieving impressive results even when only a small portion of the dataset is labeled.
We address the challenge of single-image de-raining, a task that involves recovering rain-free background information from a single rain image. While recent advancements have utilized real-world time-lapse data for training, enabling the estimation of consistent backgrounds and realistic rain streaks, these methods often suffer from computational and memory consumption, limiting their applicability in real-world scenarios. In this paper, we introduce a novel solution: the Rain Streak Prototype Unit (RsPU). The RsPU efficiently encodes rain streak-relevant features as real-time prototypes derived from time-lapse data, eliminating the need for excessive memory resources. Our de-raining network combines encoder-decoder networks with the RsPU, allowing us to learn and encapsulate diverse rain streak-relevant features as concise prototypes, employing an attention-based approach. To ensure the effectiveness of our approach, we propose a feature prototype loss encompassing cohesion and divergence components. This loss function captures both the compactness and diversity aspects of the prototypical rain streak features within the RsPU. Our method evaluates various de-raining benchmarks, accompanied by comprehensive ablation studies. We show that it can achieve competitive results in various rain images compared to state-of-the-art methods.
Previous real-time MRI (rtMRI)-based speech synthesis models depend heavily on noisy ground-truth speech. Applying loss directly over ground truth mel-spectrograms entangles speech content with MRI noise, resulting in poor intelligibility. We introduce a novel approach that adapts the multi-modal self-supervised AV-HuBERT model for text prediction from rtMRI and incorporates a new flow-based duration predictor for speaker-specific alignment. The predicted text and durations are then used by a speech decoder to synthesize aligned speech in any novel voice. We conduct thorough experiments on two datasets and demonstrate our method's generalization ability to unseen speakers. We assess our framework's performance by masking parts of the rtMRI video to evaluate the impact of different articulators on text prediction. Our method achieves a $15.18\%$ Word Error Rate (WER) on the USC-TIMIT MRI corpus, marking a huge improvement over the current state-of-the-art. Speech samples are available at https://mri2speech.github.io/MRI2Speech/