Talking head generation is an essential task in various real-world applications such as film making and virtual reality. To this end, recent works focus on the NeRF-based methods that can capture the 3D structural information of faces and generate more natural and vivid talking videos. However, the existing NeRF-based methods fail to accurately generate the audio-synced videos. In this paper, we point out that the previous methods do not consider the audio-visual representations explicitly, which is crucial for precise lip synchronization. Moreover, the existing methods struggle to generate high-frequency details, making the generation results unnatural. To overcome these problems, we propose a novel audio-synced and high-fidelity NeRF-based talking head generation framework, named Wav2NeRF, which learns audio-visual cross-modality representations and employs the wavelet transform for better visual quality. In precise, we adopt a 2D CNN-based neural rendering decoder to a NeRF-based encoder for fast generation of the whole image to employ a new multi-level SyncNet loss for accurate lip synchronization. We also propose a novel cross-attention module to effectively fuse the image and the audio representation. In addition, we integrate the wavelet transform into our framework by proposing the wavelet loss function to enhance high-frequency details. We demonstrate that the proposed method renders realistic and audio-synced talking head videos and shows outstanding performances on average in 4 representative metrics, including PSNR (+ 4.7%), SSIM (+ 2.2%), LMD (+ 51.3%), and SyncNet Confidence (+ 154.7%) compared to the NeRF-based current state-of-the-art methods.

Online action detection, which aims to identify an ongoing action from a streaming video, is an important subject in real-world applications. For this task, previous methods use recurrent neural networks for modeling temporal relations in an input sequence. However, these methods overlook the fact that the input image sequence includes not only the action of interest but background and irrelevant actions. This would induce recurrent units to accumulate unnecessary information for encoding features on the action of interest. To overcome this problem, we propose a novel recurrent unit, named Information Discrimination Unit (IDU), which explicitly discriminates the information relevancy between an ongoing action and others to decide whether to accumulate the input information. This enables learning more discriminative representations for identifying an ongoing action. In this paper, we further present a new recurrent unit, called Information Integration Unit (IIU), for action anticipation. Our IIU exploits the outputs from IDN as pseudo action labels as well as RGB frames to learn enriched features of observed actions effectively. In experiments on TVSeries and THUMOS-14, the proposed methods outperform state-of-the-art methods by a significant margin in online action detection and action anticipation. Moreover, we demonstrate the effectiveness of the proposed units by conducting comprehensive ablation studies.

Domain adaptation (DA) is a representation learning methodology that transfers knowledge from a label-sufficient source domain to a label-scarce target domain. While most of early methods are focused on unsupervised DA (UDA), several studies on semi-supervised DA (SSDA) are recently suggested. In SSDA, a small number of labeled target images are given for training, and the effectiveness of those data is demonstrated by the previous studies. However, the previous SSDA approaches solely adopt those data for embedding ordinary supervised losses, overlooking the potential usefulness of the few yet informative clues. Based on this observation, in this paper, we propose a novel method that further exploits the labeled target images for SSDA. Specifically, we utilize labeled target images to selectively generate pseudo labels for unlabeled target images. In addition, based on the observation that pseudo labels are inevitably noisy, we apply a label noise-robust learning scheme, which progressively updates the network and the set of pseudo labels by turns. Extensive experimental results show that our proposed method outperforms other previous state-of-the-art SSDA methods.

Weakly supervised object localization has recently attracted attention since it aims to identify both class labels and locations of objects by using image-level labels. Most previous methods utilize the activation map corresponding to the highest activation source. Exploiting only one activation map of the highest probability class is often biased into limited regions or sometimes even highlights background regions. To resolve these limitations, we propose to use activation maps, named combinational class activation maps (CCAM), which are linear combinations of activation maps from the highest to the lowest probability class. By using CCAM for localization, we suppress background regions to help highlighting foreground objects more accurately. In addition, we design the network architecture to consider spatial relationships for localizing relevant object regions. Specifically, we integrate non-local modules into an existing base network at both low- and high-level layers. Our final model, named non-local combinational class activation maps (NL-CCAM), obtains superior performance compared to previous methods on representative object localization benchmarks including ILSVRC 2016 and CUB- 200-2011. Furthermore, we show that the proposed method has a great capability of generalization by visualizing other datasets.

In this paper, we present a novel grid encoding model for content-aware image retargeting. In contrast to previous approaches such as vertex-based and axis-aligned grid encoding models, our approach takes each horizontal/vertical distance between two adjacent vertices as an optimization variable. Upon this difference-based encoding scheme, every vertex position of a target grid is subsequently determined after optimizing the one-dimensional values. Our quad edge-based grid model has two major advantages for image retargeting. First, the model enables a grid optimization problem to be developed in a simple quadratic program while ensuring the global convexity of objective functions. Second, due to the independency of variables, spatial regularizations can be applied in a locally adaptive manner to preserve structural components. Based on this model, we propose three quadratic objective functions. Note that, in our work, their linear combination guides a grid deformation process to obtain a visually comfortable retargeting result by preserving salient regions and structural components of an input image. Comparative evaluations have been conducted with ten existing state-of-the-art image retargeting methods, and the results show that our method built upon the quad edge-based model consistently outperforms other previous methods both on qualitative and quantitative perspectives.