WACV 2023 Papers

Learning an effective outfit-level representation is critical for predicting the compatibility of items in an outfit, and retrieving complementary items for a partial outfit. We present a framework, OutfitTransformer, that uses the proposed task-specific tokens and leverages the self-attention mechanism to learn effective outfit-level representations encoding the compatibility relations between all items in the entire outfit for addressing both compatibility prediction and complementary item retrieval. For compatibility prediction, we design an outfit token to capture a global outfit representation and train the framework using a classification loss. For complementary item retrieval, we design a target item token that additionally takes the target item specification (in the form of a category or text description) into consideration. We train our framework using a proposed set-wise outfit ranking loss to generate a target item embedding given an outfit, and a target item specification as inputs. The generated target item embedding is then used to retrieve compatible items that match the rest of the outfit. Additionally, we adopt a pre-training approach and a curriculum learning strategy to improve retrieval performance. Experiments show that our approach outperforms state-of-the-art methods on compatibility prediction, fill-in-the-blank, and complementary item retrieval tasks.

Modern deep generative models can assign high likelihood to inputs drawn from outside the training distribution, posing threats to models in open-world deployments. While much research attention has been placed on defining new test-time measures of OOD uncertainty, these methods do not fundamentally change how deep generative models are regularized and optimized in training. In particular, generative models are shown to overly rely on the background information to estimate the likelihood. To address the issue, we propose a novel frequency-regularized learning (FRL) framework for OOD detection, which incorporates high-frequency information into training and guides the model to focus on semantically relevant features. FRL effectively improves performance on a wide range of generative architectures, including variational auto-encoder, GLOW, and PixelCNN++. On a new large-scale evaluation task, FRL achieves the state-of-the-art performance, outperforming a strong baseline Likelihood Regret by 10.7% (AUROC) while achieving 147x faster inference speed. Extensive ablations show that FRL improves the OOD detection performance while preserving the image generation quality.

The task of out-of-distribution (OOD) detection is vital to realize safe and reliable operation for real-world applications. After the failure of likelihood-based detection in high dimensions had been revealed, approaches based on the typical set have been attracting attention; however, they still have not achieved satisfactory performance. Beginning by presenting the failure case of the typicality-based approach, we propose a new reconstruction error-based approach that employs normalizing flow (NF). We further introduce a typicality-based penalty, and by incorporating it into the reconstruction error in NF, we propose a new OOD detection method, penalized reconstruction error (PRE). Because the PRE detects test inputs that lie off the in-distribution manifold, it also effectively detects adversarial examples. We show the effectiveness of our method through the evaluation using natural image datasets, CIFAR-10, TinyImageNet, and ILSVRC2012.

The convolutional layers of the standard convolutional neural networks (CNNs) are equivariant to translation. Recently, a new class of CNNs is introduced which is equivariant to other affine geometric transformations such as rotation and reflection by replacing the standard convolutional layer with the group convolutional layer or using the steerable filters in the convloutional layer. We propose to embed the 2D positional encoding which is invariant to rotation, reflection and translation using orthogonal polar harmonic transforms (PHTs) before flattening the feature maps for fully-connected or classification layer in the equivariant CNN architecture. We select the PHTs among several invariant transforms, as they are very efficient in performance and speed. The proposed 2D positional encoding scheme between the convolutional and fully-connected layers of the equivariant networks is shown to provide significant improvement in performance on the rotated MNIST, CIFAR-10 and CIFAR-100 datasets.

We introduce OpenEarthMap, a benchmark dataset, for global high-resolution land cover mapping. OpenEarthMap consists of 2.2 million segments of 5000 aerial and satellite images covering 97 regions from 44 countries across 6 continents, with manually annotated 8-class land cover labels at a 0.25--0.5m ground sampling distance. Semantic segmentation models trained on the OpenEarthMap generalize worldwide and can be used as off-the-shelf models in a variety of applications. We evaluate the performance of state-of-the-art methods for unsupervised domain adaptation and present challenging problem settings suitable for further technical development. We also investigate lightweight models using automated neural architecture search for limited computational resources and fast mapping. The dataset will be made publicly available.

Multi-task learning (MTL) has found wide application in computer vision tasks. It uses a common backbone network allowing shared feature computation for different tasks such as semantic segmentation, depth- and normal estimation. In many cases negative transfer, i.e. impaired performance in the target domain, causes the MTL accuracy to be lower than simply training the corresponding single-task networks. To mitigate this issue, we propose an online knowledge distillation method for MTL, where single-task networks are trained simultaneously with the MTL network, guiding the optimization process. We propose selectively training layers for each task using an adaptive feature distillation (AFD) loss with an online task weighting (OTW) scheme. This task-wise feature distillation enables the MTL network to be trained in a similar way to the single-task networks. On the NYUv2 and Cityscapes datasets we show improvements over a baseline MTL model by 6.22% and 9.19%, respectively, and better performance than recent MTL methods. We validate our design choices, including the use of the online task weighting and the adaptive feature distillation loss in ablative experiments.

Pedestrian trajectory prediction is an essential component of autonomous systems and robot navigation. Recent research has shown promising predictive performance by designing trajectory prediction networks to model a variety of motion-related features. Different from existing works, our focus is on designing a novel online adaptation framework (OATMem) to exploit the temporal similarities among trajectory samples encountered during testing to improve the prediction accuracy of any such models (i.e., predictors) without knowing the details of these predictors. Our framework consists of two novel modules: an augmented temporal observe-target memory network (ATM) and a certainty-based selector (CS). Inspired by the concept of key-value memory networks [16], we design the ATM to learn the temporal information from short-term past frames by encoding the trajectory samples of past pedestrians in form of observe-target (i.e., key-value) during testing. In addition, we propose a certainty-based selector (CS) to enhance the predictive ability of our framework under scenarios where there are large temporal dissimilarities between current pedestrians' movements and those stored in memory. In dynamic scenes, these scenarios commonly occur due to abrupt changes in contexts, such as camera motions, scene contexts, and pedestrians' behaviors. We extensively evaluate our framework in commonly-used datasets for pedestrian trajectory prediction: JAAD [12] and PIE [19] and show that our framework significantly improves the prediction accuracy of state-of-the-art models. Finally, in-depth ablation studies and analyses are conducted to show on the importance of each proposed component.

Joint synthesis of images and segmentation masks with generative adversarial networks (GANs) is promising to reduce the effort needed for collecting image data with pixel-wise annotations. However, to learn high-fidelity image-mask synthesis, existing GAN approaches first need a pre-training phase requiring large amounts of image data, which limits their utilization in restricted image domains. In this work, we take a step to reduce this limitation, introducing the task of one-shot image-mask synthesis. We aim to generate diverse images and their segmentation masks given only a single labelled example, and assuming, contrary to previous models, no access to any pre-training data. To this end, inspired by the recent architectural developments of single-image GANs, we introduce our OSMIS model which enables the synthesis of segmentation masks that are precisely aligned to the generated images in the one-shot regime. Besides achieving the high fidelity of generated masks, OSMIS outperforms state-of-the-art single-image GAN models in image synthesis quality and diversity. In addition, despite not using any additional data, OSMIS demonstrates an impressive ability to serve as a source of useful data augmentation for one-shot segmentation applications, providing performance gains that are complementary to standard data augmentation techniques. Code is available at https://github.com/boschresearch/one-shot-synthesis.

Active consumption of digital documents has yielded scope for research in various applications, including search. Traditionally, searching within a document has been cast as a text matching problem ignoring the rich layout and visual cues commonly present in structured documents, forms, etc. To that end, we ask a mostly unexplored question: "Can we search for other similar snippets present in a target document page given a single query instance of a document snippet?". We propose MONOMER to solve this as a one-shot snippet detection task. MONOMER fuses context from visual, textual, and spatial modalities of snippets and documents to find query snippet in target documents. We conduct extensive ablations and experiments showing MONOMER outperforms several baselines from one-shot object detection (BHRL), template matching, and document understanding (LayoutLMv3). Due to the scarcity of relevant data for the task at hand, we train MONOMER on programmatically generated data having many visually similar query snippets and target document pairs from two datasets - Flamingo Forms and PubLayNet. We also do a human study to validate the generated data.

Image noise is ubiquitous in photography. However, image noise is not compressible nor desirable, thus attempting to convey the noise in compressed image bitstreams yields sub-par results in both rate and distortion. We propose to explicitly learn the image denoising task when training the codec. Therefore, we leverage the Natural Image Noise Dataset, which offers a wide variety of scenes captured with various noise levels. Given this training set, we show that a single model trained based on a mixture of images with variable noise levels appears to yield best-in-class results with both noisy and clean images, achieving better rate-distortion than a compression-only model or even than a pair of denoising-then-compression models with almost one order of magnitude fewer GMac operations.

The role of quantization within implicit/coordinate neural networks is still not fully understood. We note that using a canonical fixed quantization scheme during training produces poor performance at low bit-rates due to the network weight distributions changing over the course of training. In this work, we show that a non-uniform quantization of neural weights can lead to significant improvements. Specifically, we demonstrate that a clustered quantization enables improved reconstruction. Finally, by characterising a trade-off between quantization and network capacity, we demonstrate that it is possible (while memory inefficient) to reconstruct signals using binary neural networks. We demonstrate our findings experimentally on 2D image reconstruction and 3D radiance fields; and show that simple quantization methods and architecture search can achieve compression of NeRF to less than 16kb with minimal loss in performance (323x smaller than the original NeRF).

Synthetic image generation has recently experienced significant improvements in domains such as natural image or art generation. However, the problem of figure and diagram generation remains unexplored. A challenging aspect of generating figures and diagrams is effectively rendering readable texts within the images. To alleviate this problem, we present OCR-VQGAN, an image encoder, and decoder that leverages OCR pre-trained features to optimize a text perceptual loss, encouraging the architecture to preserve high-fidelity text and diagram structure. To explore our approach, we introduce the Paper2Fig100k dataset, with over 100k images of figures and texts from research papers. The figures show architecture diagrams and methodologies of articles available at arXiv.org from fields like artificial intelligence and computer vision. Figures usually include text and discrete objects, e.g., boxes in a diagram, with lines and arrows that connect them. We demonstrate the effectiveness of OCR-VQGAN by conducting several experiments on the task of figure reconstruction. Additionally, we explore the qualitative and quantitative impact of weighting different perceptual metrics in the overall loss function. We release code, models, and dataset at https://github.com/joanrod/ocr-vqgan.

Weakly supervised Video Anomaly Detection (wVAD) aims to distinguish anomalies from normal events based on video-level supervision. Most existing works utilize Multiple Instance Learning (MIL) with ranking loss to tackle this task. These methods, however, rely on noisy predictions from a MIL-based classifier for target instance selection in ranking loss, degrading model performance. To overcome this problem, we propose Normality Guided Multiple Instance Learning (NG-MIL) framework, which encodes diverse normal patterns from noise-free normal videos into prototypes for constructing a similarity-based classifier. By ensembling predictions of two classifiers, our method could refine the anomaly scores, reducing training instability from weak labels. Moreover, we introduce normality clustering and normality guided triplet loss constraining inner bag instances to boost the effect of NG-MIL and increase the discriminability of classifiers. Extensive experiments on three public datasets (ShanghaiTech, UCF-Crime, XD-Violence) demonstrate that our method is comparable to or better than existing weakly supervised methods, achieving state-of-the-art results.

Unsupervised Anomaly detection (AD) requires building a notion of normalcy, distinguishing in-distribution (ID) and out-of-distribution (OOD) data, using only available ID samples. Recently, large gains were made on this task for the domain of natural images using self-supervised contrastive feature learning as a first step followed by kNN or traditional one-class classifiers for feature scoring. Learned representations that are non-uniformly distributed on the unit hypersphere have been shown to be beneficial for this task. We go a step further and investigate how the ph geometrical compactness of the ID feature distribution makes isolating and detecting outliers easier, especially in the realistic situation when ID training data is polluted with OOD data. We propose novel architectural modifications to the self-supervised feature learning step, that enable such compact distributions for ID data to be learned. We show that the proposed modifications can be effectively applied to most existing self-supervised objectives, with large gains in performance. Furthermore, this improved OOD performance is obtained without resorting to tricks such as using strongly augmented ID images (e.g. by 90 degree rotations) as proxies for the unseen OOD data, as these impose overly prescriptive assumptions about ID data and its invariances. We perform extensive studies on benchmark datasets for one-class OOD detection and show state-of-the-art performance in the presence of pollution in the ID data, and comparable performance otherwise. We also propose and extensively evaluate a novel feature scoring technique based on the angular Mahalanobis distance, and propose a simple and novel technique for feature ensembling during evaluation that enables a big boost in performance at nearly zero run-time cost compared to the standard use of model ensembling or test time augmentations. Source code is available here: https://github.com/IntuitionMachines/NSA

The quality assessment (QA) of camera captured authentically distorted images is important on account of its ubiquitous applications and challenging due to the lack of a reference. While there exists a plethora of supervised no reference (NR) image QA (IQA) algorithms, there is a need to study unsupervised or opinion unaware algorithms on account of their superior generalization performance. We explore self-supervised learning (SSL) for the feature design on authentically distorted images to predict quality without training on human labels. While SSL on synthetic distortions has recently shown promise, there is a need to enrich the feature learning on authentic distortions. The key challenge in achieving this is in the learning of quality sensitive features with mitigated content dependence. We design a self-supervised contrastive learning approach which only requires positives and introduce a content separation loss by estimating a bound on the mutual information between the features learnt and the content information. We show on multiple authentically distorted datasets that our self-supervised features can predict image quality by comparing with a corpus of pristine images and achieve state-of-the-art performance.

We introduce a method for instance proposal generation for 3D point clouds. Existing techniques typically directly regress proposals in a single feed-forward step, leading to inaccurate estimation. We show that this serves as a critical bottleneck, and propose a method based on iterative bilateral filtering with learned kernels. Following the spirit of bilateral filtering, we consider both the deep feature embeddings of each point, as well as their locations in the 3D space. We show via synthetic experiments that our method brings drastic improvements when generating instance proposals for a given point of interest. We further validate our method on the challenging ScanNet benchmark, achieving the best instance segmentation performance amongst the sub-category of top-down methods.

Task incremental learning aims to enable a system to maintain its performance on previously learned tasks while learning new tasks, solving the problem of catastrophic forgetting. One promising approach is to build an individual network or sub-network for future tasks. However, this leads to an ever-growing memory due to saving extra weights for new tasks and how to address this issue has remained an open problem in task incremental learning. In this paper, we introduce a novel Neural Weight Search technique that designs a fixed search space where the optimal combinations of frozen weights can be searched to build new models for novel tasks in an end-to-end manner, resulting in a scalable and controllable memory growth. Extensive experiments on two benchmarks, i.e., Split-CIFAR-100 and CUB-to-Sketches, show our method achieves state-of-the-art performance with respect to both average inference accuracy and total memory cost.

Neural networks have recently been used to analyze diverse physical systems and to identify the underlying dynamics. While existing methods achieve impressive results, they are limited by their strong demand for training data and their weak generalization abilities to out-of-distribution data. To overcome these limitations, we propose to combine neural implicit representations for appearance modeling with neural ordinary differential equations (ODEs) for modelling physical phenomena to obtain a dynamic scene representation that can be identified directly from visual observations. Our proposed model combines several unique advantages: (i) Contrary to existing approaches that require large training datasets, we are able to identify physical parameters from only a single video. (ii) The use of neural implicit representations enables the processing of high-resolution videos and the synthesis of photo-realistic images. (iii) The embedded neural ODE has a known parametric form that allows for the identification of interpretable physical parameters, and (iv) long-term prediction in state space. (v) Furthermore, the photo-realistic rendering of novel scenes with modified physical parameters becomes possible.

We consider the problem of compressing an information source when a correlated one is available as side information only at the decoder side, which is a special case of the distributed source coding problem in information theory. In particular, we consider a pair of stereo images, which have overlapping fields of view, and are captured by a synchronized and calibrated pair of cameras as correlated image sources. In previously proposed methods, the encoder transforms the input image to a latent representation using a deep neural network, and compresses the quantized latent representation losslessly using entropy coding. The decoder decodes the entropy-coded quantized latent representation, and reconstructs the input image using this representation and the available side information. In the proposed method, the decoder employs a cross-attention module to align the feature maps obtained from the received latent representation of the input image and a latent representation of the side information. We argue that aligning the correlated patches in the feature maps allows better utilization of the side information. We empirically demonstrate the competitiveness of the proposed algorithm on KITTI and Cityscape datasets of stereo image pairs. Our experimental results show that the proposed architecture is able to exploit the decoder-only side information in a more efficient manner compared to previous works.