WACV 2023 Papers

Transformer is eminently suitable for auto-regressive image synthesis which predicts discrete value from the past values recursively to make up full image. Especially, combined with vector quantised latent representation, the state-of-the-art auto-regressive transformer displays realistic high-resolution images. However, sampling the latent code from discrete probability distribution makes the output unpredictable. Therefore, it requires to generate lots of diverse samples to acquire desired outputs. To alleviate the process of generating lots of samples repetitively, in this article, we propose to take a desired output, a style image, as an additional condition without re-training the transformer. To this end, our method transfers the style to a probability constraint to re-balance the prior, thereby specifying the target distribution instead of the original prior. Thus, generated samples from the re-balanced prior have similar styles to reference style. In practice, we can choose either an image or a category of images as an additional condition. In our qualitative assessment, we show that styles of majority of outputs are similar to the input style.

In Vision-and-Language Navigation (VLN), researchers typically take an image encoder pre-trained on ImageNet without fine-tuning on the environments that the agent will be trained or tested on. However, the distribution shift between the training images from ImageNet and the views in the navigation environments may render the ImageNet pre-trained image encoder suboptimal. Therefore, in this paper, we design a set of structure-encoding auxiliary tasks (SEA) that leverage the data in the navigation environments to pre-train and improve the image encoder. Specifically, we design and customize (1) 3D jigsaw, (2) traversability prediction, and (3) instance classification to pre-train the image encoder. Through rigorous ablations, our SEA pre-trained features are shown to better encode structural information of the scenes, which ImageNet pre-trained features fail to properly encode but is crucial for the target navigation task. The SEA pre-trained features can be easily plugged into existing VLN agents without any tuning. For example, on Test-Unseen environments, the VLN agents combined with our SEA pre-trained features achieve absolute success rate improvement of 12% for Speaker-Follower [14], 5% for Env-Dropout [37], and 4% for AuxRN [50].

One of the challenges for analyzing video contents (e.g., actions) is high computational cost, especially for the tasks that require processing densely sampled frames in a long video. We present a novel efficient action recognition algorithm, which allocates computational resources adaptively to individual frames depending on their relevance and significance. Specifically, our algorithm adopts LSTM-based policy modules and sequentially estimates the usefulness of each frame based on their intermediate representations. If a certain frame is unlikely to be helpful for recognizing actions, our model stops forwarding the features to the rest of the layers and starts to consider the next sampled frame. We further reduce the computational cost of our approach by introducing a simple yet effective early termination strategy during the inference procedure. We evaluate the proposed algorithm on three public benchmarks: ActivityNet-v1.3, Mini-Kinetics, and THUMOS'14. Our experiments show that the proposed approach achieves outstanding trade-off between accuracy and efficiency in action recognition.

This paper presents an approach to train a unified deep network that simultaneously solves multiple human-related tasks. A multi-task framework is favorable for sharing information across tasks under restricted computational resources. However, tasks not only share information but may also compete for resources and conflict with each other, making the optimization of shared parameters difficult and leading to suboptimal performance. We propose a simple but effective training scheme called GradSplit that alleviates this issue by utilizing asymmetric inter-task relations. Specifically, at each convolution module, it splits features into T groups for T tasks and trains each group only using the gradient back-propagated from the task losses with which it does not have conflicts. During training, we apply GradSplit to a series of convolution modules. As a result, each module is trained to generate a set of task-specific features using the shared features from the previous module. This enables a network to use complementary information across tasks while circumventing gradient conflicts. Experimental results show that GradSplit achieves a better accuracy-efficiency trade-off than existing methods. It minimizes accuracy drop caused by task conflicts while significantly saving compute resources in terms of both FLOPs and memory at inference. We further show that GradSplit achieves higher cross-dataset accuracy compared to single-task and other multi-task networks.

Frame interpolation is an essential video processing technique that adjusts the temporal resolution of an image sequence. While deep learning has brought great improvements to the area of video frame interpolation, techniques that make use of neural networks can typically not easily be deployed in practical applications like a video editor since they are either computationally too demanding or fail at high resolutions. In contrast, we propose a deep learning approach that solely relies on splatting to synthesize interpolated frames. This splatting-based synthesis for video frame interpolation is not only much faster than similar approaches, especially for multi-frame interpolation, but can also yield new state-of-the-art results at high resolutions.

In many emerging computer vision applications, it is critical to adhere to stringent latency and power constraints. The current neural network paradigm of frame-based, floating-point inference is often ill-suited to these resource-constrained applications. Spike-based perception - enabled by spiking neural networks (SNNs) - is one promising alternative. Unlike conventional neural networks (ANNs), spiking networks exhibit smooth tradeoffs between latency, power, and accuracy. SNNs are the archetype of an "anytime algorithm" whose accuracy improves smoothly over time. This property allows SNNs to adapt their computational investment in response to changing resource constraints. Unfortunately, mainstream algorithms for training SNNs (i.e., those based on ANN-to-SNN conversion) tend to produce models that are inefficient in practice. To mitigate this problem, we propose a set of principled optimizations that reduce latency and power consumption by 1-2 orders of magnitude in converted SNNs. These optimizations leverage a set of novel efficiency metrics designed for anytime algorithms. We also develop a state-of-the-art simulator, SaRNN, which can simulate SNNs using commodity GPU hardware and neuromorphic platforms. We hope that the proposed optimizations, metrics, and tools will facilitate the future development of spike-based vision systems.

Current methods for spatiotemporal action tube detection often extend a bounding box proposal at a given key-frame into a 3D temporal cuboid and pool features from nearby frames. However, such pooling fails to accumulate meaningful spatiotemporal features if the position or shape of the actor shows large 2D motion and variability through the frames, due to large camera motion, large actor shape deformation, fast actor action and so on. In this work, we aim to study the performance of cuboid-aware feature aggregation in action detection under large action. Further, we propose to enhance actor feature representation under large motion by tracking actors and performing temporal feature aggregation along the respective tracks. We define the actor motion with intersection-over-union (IoU) between the boxes of action tubes/tracks at various fixed time scales. The action having a large motion would result in lower IoU over time, and slower actions would maintain higher IoU. We find that track-aware feature aggregation consistently achieves a large improvement in action detection performance, especially for actions under large motion compared to the cuboid-aware baseline. As a result, we also report state-of-the-art on the large-scale MultiSports dataset.

In most scenarios, conditional image generation can be thought of as an inversion of the image understanding process. Since generic image understanding involves solving multiple tasks, it is natural to aim at generating images via multi conditioning. However, multi-conditional image generation is a very challenging problem due to the heterogeneity and the sparsity of the (in practice) available conditioning labels. In this work, we propose a novel neural architecture to address the problem of heterogeneity and sparsity of the spatially multi-conditional labels. Our choice of spatial conditioning, such as by semantics and depth, is driven by the promise it holds for better control of the image generation process. The proposed method uses a transformer-like architecture operating pixel-wise, which receives the available labels as input tokens to merge them in a learned homogeneous space of labels. The merged labels are then used for image generation via conditional generative adversarial training. In this process, the sparsity of the labels is handled by simply dropping the input tokens corresponding to the missing labels at the desired locations, thanks to the proposed pixel-wise operating architecture. Our experiments on three benchmark datasets demonstrate the clear superiority of our method over the state-of-the-art and compared baselines.

Multi-label image classification is more applicable 'in the wild' than single-label classification, as natural images usually contain multiple objects. However, exhaustively annotating images with every object of interest is costly and time-consuming. We train multi-label classifiers from datasets where each image is annotated with a single positive label only. As the presence of all other classes is unknown, we propose an Expected Negative loss that builds a set of expected negative labels in addition to the annotated positives. This set is determined based on prediction consistency, by averaging predictions over consecutive training epochs to build robust targets. Moreover, the crop data-augmentation leads to additional label noise by cropping out the single annotated object. Our novel spatial consistency loss improves supervision and ensures consistency of the spatial feature maps by maintaining per-class running-average heatmaps for each training image. We use MS-COCO, Pascal VOC, NUS-WIDE and CUB-Birds datasets to demonstrate the gains of the Expected Negative loss in combination with consistency and spatial consistency losses. We also demonstrate improved multi-label classification mAP on ImageNet-1K using the ReaL multi-label validation set.

We present a novel depth completion approach agnostic to the sparsity of depth points, that is very likely to vary in many practical applications. State-of-the-art approaches yield accurate results only when processing a specific density and distribution of input points, i.e. the one observed during training, narrowing their deployment in real use cases. On the contrary, our solution is robust to uneven distributions and extremely low densities never witnessed during training. Experimental results on standard indoor and outdoor benchmarks highlight the robustness of our framework, achieving accuracy comparable to state-of-the-art methods when tested with density and distribution equal to the training one while being much more accurate in the other cases. Our pretrained models and further material are available in our project page.

Humans are, arguably, one of the most important regions of interest in a visual analysis pipeline. Detecting how the human interacts with the surrounding environment, thus, becomes an important problem and has several potential use-cases. While this has been adequately addressed in the literature in the image setting, there exist very few methods addressing the case for in-the-wild videos. The problem is further exacerbated by the high degree of label skew. To this end, we propose SeRVo-HOI, a robust end-to-end framework for recognizing human-object interactions from a video, particularly in high label-skew settings. The network contextualizes multiple image representations and is trained to explicitly handle dataset skew. We propose and analyse methods to address the long-tail distribution of the labels and show improvements on the tail-labels. SeRVo-HOI outperforms the state-of-the-art by a significant margin 21.1% vs 17.6% mAP on the large-scale, in-the-wild VidHOI dataset while particularly demonstrating solid improvements in the tail-classes 19.9% vs 17.3% mAP.

This paper proposes the first GAN inversion-based method for multi-class sketch-based image generation (MC-SBIG). MC-SBIG is a challenging task that requires strong prior knowledge due to the significant domain gap between sketches and natural images. Existing learning-based approaches rely on a large-scale paired dataset to learn the mapping between these two image modalities. However, since the public paired sketch-photo data are scarce, it is struggling for learning-based methods to achieve satisfactory results. In this work, we introduce a new approach based on GAN inversion, which can utilize a powerful pretrained generator to facilitate image generation from a given sketch. Our GAN inversion-based method has two advantages: 1. it can freely take advantage of the prior knowledge of a pretrained image generator; 2. it allows the proposed model to focus on learning the mapping from a sketch to a low-dimension latent code, which is a much easier task than directly mapping to a high-dimension natural image. We also present a novel shape loss to improve generation quality further. Extensive experiments are conducted to show that our method can produce sketch-faithful and photo-realistic images and significantly outperform the baseline methods.

High dynamic range (HDR) imaging is an indispensable technique in modern photography. Traditional methods focus on HDR reconstruction from multiple images, solving the core problems of image alignment, fusion, and tone mapping, yet having a perfect solution due to ghosting and other visual artifacts in the reconstruction. Recent attempts at single-image HDR reconstruction show a promising alternative: by learning to map pixel values to their irradiance using a neural network, one can bypass the align-and-merge pipeline completely yet still obtain a high-quality HDR image. In this work, we propose a weakly supervised learning method that inverts the physical image formation process for HDR reconstruction via learning to generate multiple exposures from a single image. Our neural network can invert the camera response to reconstruct pixel irradiance before synthesizing multiple exposures and hallucinating details in under- and over-exposed regions from a single input image. To train the network, we propose a representation loss, a reconstruction loss, and a perceptual loss applied on pairs of under- and over-exposure images and thus do not require HDR images for training. Our experiments show that our proposed model can effectively reconstruct HDR images. Our qualitative and quantitative results show that our method achieves state-of-the-art performance on the DrTMO dataset. Our code is available at https://github.com/VinAIResearch/single_image_hdr.

The costly process of obtaining semantic segmentation labels has driven research towards weakly supervised semantic segmentation (WSSS) methods, using only image-level, point, or box labels. Such annotations introduce limitations and challenges that results in overly-tuned methods specialized in specific domains or scene types. The over reliance of image-level based methods on generation of high quality class activation maps (CAMs) results in limited applicable dataset complexity range, mostly focusing on object centric scenes. Additionally, the lack of dense annotations requires methods to increase network complexity to obtain additional semantic information, often done through multiple stages of training and refinement. Here, we present a single-stage approach generalizable to a wide range of dataset complexities, that is trainable from scratch, without any dependency on pre-trained backbones, classification, or separate refinement tasks. We utilize point annotations to generate reliable, on-the-fly pseudo-masks through refined and spatially filtered features. We are to demonstrate SOTA performance on benchmark datasets (PascalVOC 2012), as well as significantly outperform other SOTA WSSS methods on recent real-world datasets (CRAID, CityPersons, IAD, ADE20K, CityScapes) with up to 28.1% and 22.6% performance boosts compared to our single-stage and multi-stage baselines respectively.

In this paper, we introduce a novel implicit neural network for the task of single image super-resolution at arbitrary scale factors. To do this, we represent an image as a decoding function that maps locations in the image along with their associated features to their reciprocal pixel attributes. Since the pixel locations are continuous in this representation, our method can refer to any location in an image of varying resolution. To retrieve an image of a particular resolution, we apply a decoding function to a grid of locations each of which refers to the center of a pixel in the output image. In contrast to other techniques, our dual interactive neural network decouples content and positional features. As a result, we obtain a fully implicit representation of the image that solves the super-resolution problem at (real-valued) elective scales using a single model. We demonstrate the efficacy and flexibility of our approach against the state of the art on publicly available benchmark datasets.

This paper proposes a novel polarization sensor structure and network architecture to obtain a high-quality RGB image and polarization information. Conventional polarization sensors can simultaneously acquire RGB images and polarization information, but the polarizers on the sensor degrade the quality of the RGB images. There is a trade-off between the quality of the RGB image and polarization information as fewer polarization pixels reduce the degradation of the RGB image but decrease the resolution of polarization information. Therefore, we propose an approach that resolves the trade-off by sparsely arranging polarization pixels on the sensor and compensating for low-resolution polarization information with higher resolution using the RGB image as a guide. Our proposed network architecture consists of an RGB image refinement network and a polarization information compensation network. We confirmed the superiority of our proposed network in compensating the differential component of polarization intensity by comparing its performance with state-of-the-art methods for similar tasks: depth completion. Furthermore, we confirmed that our approach could simultaneously acquire higher quality RGB images and polarization information than conventional polarization sensors, resolving the trade-off between the quality of RGB images and polarization information. The baseline code and newly generated real and synthetic large-scale polarization image datasets are available for further research and development.

Contrastive representation learning has proven to be an effective self-supervised learning method. Most successful approaches are based on Noise Contrastive Estimation (NCE) and use different views of an instance as positives that should be contrasted with other instances, called negatives, that are considered as noise. However, several instances in a dataset are drawn from the same distribution and share underlying semantic information. A good data representation should contain relations, or semantic similarity, between the instances. Contrastive learning implicitly learns relations but considering all negatives as noise harms the quality of the learned relations. To circumvent this issue, we propose a novel formulation of contrastive learning using semantic similarity between instances called Similarity Contrastive Estimation (SCE). Our training objective is a soft contrastive learning one. Instead of hard classifying positives and negatives, we estimate from one view of a batch a continuous distribution to push or pull instances based on their semantic similarities. This target similarity distribution is sharpened to eliminate noisy relations. The model predicts for each instance, from another view, the target distribution while contrasting its positive with negatives. Experimental results show that SCE is Top-1 on the ImageNet linear evaluation protocol at 100 pretraining epochs with 72.1% accuracy and is competitive with state-of-the-art algorithms by reaching 75.4% for 200 epochs with multi-crop. We also show that SCE is able to generalize to several tasks. Source code is available here: https://github.com/CEA-LIST/SCE.

An agent with a limited field of view needs to sample the most informative local observations of an environment in order to model the global context. Current works train this selection strategy by defining a complex architecture built upon features learned through convolutional encoders. In this paper, we first discuss why vision transformers are better suited than CNNs for such an agent. Next, we propose a simple transformer based active visual sampling model, called "SimGlim", which utilises transformer's inherent self-attention architecture to sequentially predict the best next location based on the current observable environment. We show the efficacy of our proposed method on the task of image reconstruction in the partial observable setting and compare our model against existing state-of-the-art active visual reconstruction methods. Finally, we provide ablations for the parameters of our design choice to understand their importance in the overall architecture.

On robotics computer vision tasks, generating and annotating large amounts of data from real-world for the use of deep learning-based approaches is often difficult or even impossible. A common strategy for solving this problem is to apply simulation-to-reality (sim2real) approaches with the help of simulated scenes. While the majority of current robotics vision sim2real work focuses on image data, we present an industrial application case that uses sim2real transfer learning for point cloud data. We provide insights on how to generate and process synthetic point cloud data in order to achieve better performance when the learned model is transferred to real-world data. The issue of imbalanced learning is investigated using multiple strategies. A novel patch-based attention network is proposed additionally to tackle this problem.