Dynamic NeRFs have recently garnered growing attention for 3D talking portrait synthesis. Despite advances in rendering speed and visual quality, challenges persist in enhancing efficiency and effectiveness. We present R2-Talker, an efficient and effective framework enabling realistic real-time talking head synthesis. Specifically, using multi-resolution hash grids, we introduce a novel approach for encoding facial landmarks as conditional features. This approach losslessly encodes landmark structures as conditional features, decoupling input diversity, and conditional spaces by mapping arbitrary landmarks to a unified feature space. We further propose a scheme of progressive multilayer conditioning in the NeRF rendering pipeline for effective conditional feature fusion. Our new approach has the following advantages as demonstrated by extensive experiments compared with the state-of-the-art works: 1) The lossless input encoding enables acquiring more precise features, yielding superior visual quality. The decoupling of inputs and conditional spaces improves generalizability. 2) The fusing of conditional features and MLP outputs at each MLP layer enhances conditional impact, resulting in more accurate lip synthesis and better visual quality. 3) It compactly structures the fusion of conditional features, significantly enhancing computational efficiency.
Query-focused summarization (QFS) aims to provide a summary of a single document/multi documents that can satisfy the information needs of a given query. It is useful for various real-world applications, such as abstractive snippet generation or more recent retrieval augmented generation (RAG). A prototypical QFS pipeline consists of a retriever (sparse or dense retrieval) and a generator (usually a large language model). However, applying large language models (LLM) potentially leads to hallucinations, especially when the evidence contradicts the prior belief of LLMs. There has been growing interest in developing new decoding methods to improve generation quality and reduce hallucination. In this work, we conduct a large-scale reproducibility study on one recently proposed decoding method -- Context-aware Decoding (CAD). In addition to replicating CAD's experiments on news summarization datasets, we include experiments on QFS datasets, and conduct more rigorous analysis on computational complexity and hyperparameter sensitivity. Experiments with eight different language models show that performance-wise, CAD improves QFS quality by (1) reducing factuality errors/hallucinations while (2) mostly retaining the match of lexical patterns, measured by ROUGE scores, while also at a cost of increased inference-time FLOPs and reduced decoding speed. The code implementation based on Huggingface Library is made available https://github.com/zhichaoxu-shufe/context-aware-decoding-qfs
We consider the problem of tracking multiple, unknown, and time-varying numbers of objects using a distributed network of heterogeneous sensors. In an effort to derive a formulation for practical settings, we consider limited and unknown sensor field-of-views (FoVs), sensors with limited local computational resources and communication channel capacity. The resulting distributed multi-object tracking algorithm involves solving an NP-hard multidimensional assignment problem either optimally for small-size problems or sub-optimally for general practical problems. For general problems, we propose an efficient distributed multi-object tracking algorithm that performs track-to-track fusion using a clustering-based analysis of the state space transformed into a density space to mitigate the complexity of the assignment problem. The proposed algorithm can more efficiently group local track estimates for fusion than existing approaches. To ensure we achieve globally consistent identities for tracks across a network of nodes as objects move between FoVs, we develop a graph-based algorithm to achieve label consensus and minimise track segmentation. Numerical experiments with a synthetic and a real-world trajectory dataset demonstrate that our proposed method is significantly more computationally efficient than state-of-the-art solutions, achieving similar tracking accuracy and bandwidth requirements but with improved label consistency.
Brain tumors are one of the deadliest forms of cancer with a mortality rate of over 80%. A quick and accurate diagnosis is crucial to increase the chance of survival. However, in medical analysis, the manual annotation and segmentation of a brain tumor can be a complicated task. Multiple MRI modalities are typically analyzed as they provide unique information regarding the tumor regions. Although these MRI modalities are helpful for segmenting gliomas, they tend to increase overfitting and computation. This paper proposes a region of interest detection algorithm that is implemented during data preprocessing to locate salient features and remove extraneous MRI data. This decreases the input size, allowing for more aggressive data augmentations and deeper neural networks. Following the preprocessing of the MRI modalities, a fully convolutional autoencoder with soft attention segments the different brain MRIs. When these deep learning algorithms are implemented in practice, analysts and physicians cannot differentiate between accurate and inaccurate predictions. Subsequently, test time augmentations and an energy-based model were used for voxel-based uncertainty predictions. Experimentation was conducted on the BraTS benchmarks and achieved state-of-the-art segmentation performance. Additionally, qualitative results were used to assess the segmentation models and uncertainty predictions.
Speech emotion recognition (SER) has received a great deal of attention in recent years in the context of spontaneous conversations. While there have been notable results on datasets like the well known corpus of naturalistic dyadic conversations, IEMOCAP, for both the case of categorical and dimensional emotions, there are few papers which try to predict both paradigms at the same time. Therefore, in this work, we aim to highlight the performance contribution of multi-task learning by proposing a multi-task, multi-modal system that predicts categorical and dimensional emotions. The results emphasise the importance of cross-regularisation between the two types of emotions. Our approach consists of a multi-task, multi-modal architecture that uses parallel feature refinement through self-attention for the feature of each modality. In order to fuse the features, our model introduces a set of learnable bridge tokens that merge the acoustic and linguistic features with the help of cross-attention. Our experiments for categorical emotions on 10-fold validation yield results comparable to the current state-of-the-art. In our configuration, our multi-task approach provides better results compared to learning each paradigm separately. On top of that, our best performing model achieves a high result for valence compared to the previous multi-task experiments.
Within recent approaches to text-to-video (T2V) generation, achieving controllability in the synthesized video is often a challenge. Typically, this issue is addressed by providing low-level per-frame guidance in the form of edge maps, depth maps, or an existing video to be altered. However, the process of obtaining such guidance can be labor-intensive. This paper focuses on enhancing controllability in video synthesis by employing straightforward bounding boxes to guide the subject in various ways, all without the need for neural network training, finetuning, optimization at inference time, or the use of pre-existing videos. Our algorithm, TrailBlazer, is constructed upon a pre-trained (T2V) model, and easy to implement. The subject is directed by a bounding box through the proposed spatial and temporal attention map editing. Moreover, we introduce the concept of keyframing, allowing the subject trajectory and overall appearance to be guided by both a moving bounding box and corresponding prompts, without the need to provide a detailed mask. The method is efficient, with negligible additional computation relative to the underlying pre-trained model. Despite the simplicity of the bounding box guidance, the resulting motion is surprisingly natural, with emergent effects including perspective and movement toward the virtual camera as the box size increases.
This work proposes TimeChat, a time-sensitive multimodal large language model specifically designed for long video understanding. Our model incorporates two key architectural contributions: (1) a timestamp-aware frame encoder that binds visual content with the timestamp of each frame, and (2) a sliding video Q-Former that produces a video token sequence of varying lengths to accommodate videos of various durations. Additionally, we construct an instruction-tuning dataset, encompassing 6 tasks and a total of 125K instances, to further enhance TimeChat's instruction-following performance. Experiment results across various video understanding tasks, such as dense captioning, temporal grounding, and highlight detection, demonstrate TimeChat's strong zero-shot temporal localization and reasoning capabilities. For example, it achieves +9.2 F1 score and +2.8 CIDEr on YouCook2, +5.8 HIT@1 on QVHighlights, and +27.5 R@1 (IoU=0.5) on Charades-STA, compared to state-of-the-art video large language models, holding the potential to serve as a versatile video assistant for long-form video comprehension tasks and satisfy realistic user requirements.
Recently, researchers have attempted to investigate the capability of LLMs in handling videos and proposed several video LLM models. However, the ability of LLMs to handle video grounding (VG), which is an important time-related video task requiring the model to precisely locate the start and end timestamps of temporal moments in videos that match the given textual queries, still remains unclear and unexplored in literature. To fill the gap, in this paper, we propose the LLM4VG benchmark, which systematically evaluates the performance of different LLMs on video grounding tasks. Based on our proposed LLM4VG, we design extensive experiments to examine two groups of video LLM models on video grounding: (i) the video LLMs trained on the text-video pairs (denoted as VidLLM), and (ii) the LLMs combined with pretrained visual description models such as the video/image captioning model. We propose prompt methods to integrate the instruction of VG and description from different kinds of generators, including caption-based generators for direct visual description and VQA-based generators for information enhancement. We also provide comprehensive comparisons of various VidLLMs and explore the influence of different choices of visual models, LLMs, prompt designs, etc, as well. Our experimental evaluations lead to two conclusions: (i) the existing VidLLMs are still far away from achieving satisfactory video grounding performance, and more time-related video tasks should be included to further fine-tune these models, and (ii) the combination of LLMs and visual models shows preliminary abilities for video grounding with considerable potential for improvement by resorting to more reliable models and further guidance of prompt instructions.
In this paper, we propose an efficient and accurate streaming speech recognition model based on the FastConformer architecture. We adapted the FastConformer architecture for streaming applications through: (1) constraining both the look-ahead and past contexts in the encoder, and (2) introducing an activation caching mechanism to enable the non-autoregressive encoder to operate autoregressively during inference. The proposed model is thoughtfully designed in a way to eliminate the accuracy disparity between the train and inference time which is common for many streaming models. Furthermore, our proposed encoder works with various decoder configurations including Connectionist Temporal Classification (CTC) and RNN-Transducer (RNNT) decoders. Additionally, we introduced a hybrid CTC/RNNT architecture which utilizes a shared encoder with both a CTC and RNNT decoder to boost the accuracy and save computation. We evaluate the proposed model on LibriSpeech dataset and a multi-domain large scale dataset and demonstrate that it can achieve better accuracy with lower latency and inference time compared to a conventional buffered streaming model baseline. We also showed that training a model with multiple latencies can achieve better accuracy than single latency models while it enables us to support multiple latencies with a single model. Our experiments also showed the hybrid architecture would not only speedup the convergence of the CTC decoder but also improves the accuracy of streaming models compared to single decoder models.
Manual delineation of tumor regions from magnetic resonance (MR) images is time-consuming, requires an expert, and is prone to human error. In recent years, deep learning models have been the go-to approach for the segmentation of brain tumors. U-Net and its' variants for semantic segmentation of medical images have achieved good results in the literature. However, U-Net and its' variants tend to over-segment tumor regions and may not accurately segment the tumor edges. The edges of the tumor are as important as the tumor regions for accurate diagnosis, surgical precision, and treatment planning. In the proposed work, the authors aim to extract edges from the ground truth using a derivative-like filter followed by edge reconstruction to obtain an edge ground truth in addition to the brain tumor ground truth. Utilizing both ground truths, the author studies several U-Net and its' variant architectures with and without tumor edges ground truth as a target along with the tumor ground truth for brain tumor segmentation. The author used the BraTS2020 benchmark dataset to perform the study and the results are tabulated for the dice and Hausdorff95 metrics. The mean and median metrics are calculated for the whole tumor (WT), tumor core (TC), and enhancing tumor (ET) regions. Compared to the baseline U-Net and its variants, the models that learned edges along with the tumor regions performed well in core tumor regions in both training and validation datasets. The improved performance of edge-trained models trained on baseline models like U-Net and V-Net achieved performance similar to baseline state-of-the-art models like Swin U-Net and hybrid MR-U-Net. The edge-target trained models are capable of generating edge maps that can be useful for treatment planning. Additionally, for further explainability of the results, the activation map generated by the hybrid MR-U-Net has been studied.