To avoid collapse in self-supervised learning (SSL), a contrastive loss is widely used but often requires a large number of negative samples. Without negative samples yet achieving competitive performance, a recent work has attracted significant attention for providing a minimalist simple Siamese (SimSiam) method to avoid collapse. However, the reason for how it avoids collapse without negative samples remains not fully clear and our investigation starts by revisiting the explanatory claims in the original SimSiam. After refuting their claims, we introduce vector decomposition for analyzing the collapse based on the gradient analysis of the $l_2$-normalized representation vector. This yields a unified perspective on how negative samples and SimSiam alleviate collapse. Such a unified perspective comes timely for understanding the recent progress in SSL.
Video frame interpolation (VFI) works generally predict intermediate frame(s) by first estimating the motion between inputs and then warping the inputs to the target time with the estimated motion. This approach, however, is not optimal when the temporal distance between the input sequence increases as existing motion estimation modules cannot effectively handle large motions. Hence, VFI works perform well for small frame gaps and perform poorly as the frame gap increases. In this work, we propose a novel framework to address this problem. We argue that when there is a large gap between inputs, instead of estimating imprecise motion that will eventually lead to inaccurate interpolation, we can safely propagate from one side of the input up to a reliable time frame using the other input as a reference. Then, the rest of the intermediate frames can be interpolated using standard approaches as the temporal gap is now narrowed. To this end, we propose a propagation network (PNet) by extending the classic feature-level forecasting with a novel motion-to-feature approach. To be thorough, we adopt a simple interpolation model along with PNet as our full model and design a simple procedure to train the full model in an end-to-end manner. Experimental results on several benchmark datasets confirm the effectiveness of our method for long-term VFI compared to state-of-the-art approaches.
Human brain is continuously inundated with the multisensory information and their complex interactions coming from the outside world at any given moment. Such information is automatically analyzed by binding or segregating in our brain. While this task might seem effortless for human brains, it is extremely challenging to build a machine that can perform similar tasks since complex interactions cannot be dealt with single type of integration but requires more sophisticated approaches. In this paper, we propose a new model to address the multisensory integration problem with individual event-specific layers in a multi-task learning scheme. Unlike previous works where single type of fusion is used, we design event-specific layers to deal with different audio-visual relationship tasks, enabling different ways of audio-visual formation. Experimental results show that our event-specific layers can discover unique properties of the audio-visual relationships in the videos. Moreover, although our network is formulated with single labels, it can output additional true multi-labels to represent the given videos. We demonstrate that our proposed framework also exposes the modality bias of the video data category-wise and dataset-wise manner in popular benchmark datasets.
Adversarial training (AT) and its variants are the most effective approaches for obtaining adversarially robust models. A unique characteristic of AT is that an inner maximization problem needs to be solved repeatedly before the model weights can be updated, which makes the training slow. FGSM AT significantly improves its efficiency but it fails when the step size grows. The SOTA GradAlign makes FGSM AT compatible with a higher step size, however, its regularization on input gradient makes it 3 to 4 times slower than FGSM AT. Our proposed NoiseAug removes the extra computation overhead by directly regularizing on the input itself. The key contribution of this work lies in an empirical finding that single-step FGSM AT is not as hard as suggested in the past line of work: noise augmentation is all you need for (FGSM) fast AT. Towards understanding the success of our NoiseAug, we perform an extensive analysis and find that mitigating Catastrophic Overfitting (CO) and Robust Overfitting (RO) need different augmentations. Instead of more samples caused by data augmentation, we identify what makes NoiseAug effective for preventing CO might lie in its improved local linearity.
The objective of this work is to localize the sound sources in visual scenes. Existing audio-visual works employ contrastive learning by assigning corresponding audio-visual pairs from the same source as positives while randomly mismatched pairs as negatives. However, these negative pairs may contain semantically matched audio-visual information. Thus, these semantically correlated pairs, "hard positives", are mistakenly grouped as negatives. Our key contribution is showing that hard positives can give similar response maps to the corresponding pairs. Our approach incorporates these hard positives by adding their response maps into a contrastive learning objective directly. We demonstrate the effectiveness of our approach on VGG-SS and SoundNet-Flickr test sets, showing favorable performance to the state-of-the-art methods.
Recently, self-supervised learning of depth and ego-motion from thermal images shows strong robustness and reliability under challenging scenarios. However, the inherent thermal image properties such as weak contrast, blurry edges, and noise hinder to generate effective self-supervision from thermal images. Therefore, most research relies on additional self-supervision sources such as well-lit RGB images, generative models, and Lidar information. In this paper, we conduct an in-depth analysis of thermal image characteristics that degenerates self-supervision from thermal images. Based on the analysis, we propose an effective thermal image mapping method that significantly increases image information, such as overall structure, contrast, and details, while preserving temporal consistency. The proposed method shows outperformed depth and pose results than previous state-of-the-art networks without leveraging additional RGB guidance.
MLP-Mixer has newly appeared as a new challenger against the realm of CNNs and transformer. Despite its simplicity compared to transformer, the concept of channel-mixing MLPs and token-mixing MLPs achieves noticeable performance in visual recognition tasks. Unlike images, point clouds are inherently sparse, unordered and irregular, which limits the direct use of MLP-Mixer for point cloud understanding. In this paper, we propose PointMixer, a universal point set operator that facilitates information sharing among unstructured 3D points. By simply replacing token-mixing MLPs with a softmax function, PointMixer can "mix" features within/between point sets. By doing so, PointMixer can be broadly used in the network as inter-set mixing, intra-set mixing, and pyramid mixing. Extensive experiments show the competitive or superior performance of PointMixer in semantic segmentation, classification, and point reconstruction against transformer-based methods.
Many mobile manufacturers recently have adopted Dual-Pixel (DP) sensors in their flagship models for faster auto-focus and aesthetic image captures. Despite their advantages, research on their usage for 3D facial understanding has been limited due to the lack of datasets and algorithmic designs that exploit parallax in DP images. This is because the baseline of sub-aperture images is extremely narrow and parallax exists in the defocus blur region. In this paper, we introduce a DP-oriented Depth/Normal network that reconstructs the 3D facial geometry. For this purpose, we collect a DP facial data with more than 135K images for 101 persons captured with our multi-camera structured light systems. It contains the corresponding ground-truth 3D models including depth map and surface normal in metric scale. Our dataset allows the proposed matching network to be generalized for 3D facial depth/normal estimation. The proposed network consists of two novel modules: Adaptive Sampling Module and Adaptive Normal Module, which are specialized in handling the defocus blur in DP images. Finally, the proposed method achieves state-of-the-art performances over recent DP-based depth/normal estimation methods. We also demonstrate the applicability of the estimated depth/normal to face spoofing and relighting.
Learning to estimate object pose often requires ground-truth (GT) labels, such as CAD model and absolute-scale object pose, which is expensive and laborious to obtain in the real world. To tackle this problem, we propose an unsupervised domain adaptation (UDA) for category-level object pose estimation, called \textbf{UDA-COPE}. Inspired by the recent multi-modal UDA techniques, the proposed method exploits a teacher-student self-supervised learning scheme to train a pose estimation network without using target domain labels. We also introduce a bidirectional filtering method between predicted normalized object coordinate space (NOCS) map and observed point cloud, to not only make our teacher network more robust to the target domain but also to provide more reliable pseudo labels for the student network training. Extensive experimental results demonstrate the effectiveness of our proposed method both quantitatively and qualitatively. Notably, without leveraging target-domain GT labels, our proposed method achieves comparable or sometimes superior performance to existing methods that depend on the GT labels.
Point cloud obtained from 3D scanning is often sparse, noisy, and irregular. To cope with these issues, recent studies have been separately conducted to densify, denoise, and complete inaccurate point cloud. In this paper, we advocate that jointly solving these tasks leads to significant improvement for point cloud reconstruction. To this end, we propose a deep point cloud reconstruction network consisting of two stages: 1) a 3D sparse stacked-hourglass network as for the initial densification and denoising, 2) a refinement via transformers converting the discrete voxels into 3D points. In particular, we further improve the performance of transformer by a newly proposed module called amplified positional encoding. This module has been designed to differently amplify the magnitude of positional encoding vectors based on the points' distances for adaptive refinements. Extensive experiments demonstrate that our network achieves state-of-the-art performance among the recent studies in the ScanNet, ICL-NUIM, and ShapeNetPart datasets. Moreover, we underline the ability of our network to generalize toward real-world and unmet scenes.