We propose an approach to semantic segmentation that achieves state-of-the-art supervised performance when applied in a zero-shot setting. It thus achieves results equivalent to those of the supervised methods, on each of the major semantic segmentation datasets, without training on those datasets. This is achieved by replacing each class label with a vector-valued embedding of a short paragraph that describes the class. The generality and simplicity of this approach enables merging multiple datasets from different domains, each with varying class labels and semantics. The resulting merged semantic segmentation dataset of over 2 Million images enables training a model that achieves performance equal to that of state-of-the-art supervised methods on 7 benchmark datasets, despite not using any images therefrom. By fine-tuning the model on standard semantic segmentation datasets, we also achieve a significant improvement over the state-of-the-art supervised segmentation on NYUD-V2 and PASCAL-context at 60% and 65% mIoU, respectively. Based on the closeness of language embeddings, our method can even segment unseen labels. Extensive experiments demonstrate strong generalization to unseen image domains and unseen labels, and that the method enables impressive performance improvements in downstream applications, including depth estimation and instance segmentation.
We propose a direct, regression-based approach to 2D human pose estimation from single images. We formulate the problem as a sequence prediction task, which we solve using a Transformer network. This network directly learns a regression mapping from images to the keypoint coordinates, without resorting to intermediate representations such as heatmaps. This approach avoids much of the complexity associated with heatmap-based approaches. To overcome the feature misalignment issues of previous regression-based methods, we propose an attention mechanism that adaptively attends to the features that are most relevant to the target keypoints, considerably improving the accuracy. Importantly, our framework is end-to-end differentiable, and naturally learns to exploit the dependencies between keypoints. Experiments on MS-COCO and MPII, two predominant pose-estimation datasets, demonstrate that our method significantly improves upon the state-of-the-art in regression-based pose estimation. More notably, ours is the first regression-based approach to perform favorably compared to the best heatmap-based pose estimation methods.
Graph-level anomaly detection (GAD) describes the problem of detecting graphs that are abnormal in their structure and/or the features of their nodes, as compared to other graphs. One of the challenges in GAD is to devise graph representations that enable the detection of both locally- and globally-anomalous graphs, i.e., graphs that are abnormal in their fine-grained (node-level) or holistic (graph-level) properties, respectively. To tackle this challenge we introduce a novel deep anomaly detection approach for GAD that learns rich global and local normal pattern information by joint random distillation of graph and node representations. The random distillation is achieved by training one GNN to predict another GNN with randomly initialized network weights. Extensive experiments on 16 real-world graph datasets from diverse domains show that our model significantly outperforms seven state-of-the-art models. Code and datasets are available at https://git.io/GLocalKD.
Existing anomaly detection paradigms overwhelmingly focus on training detection models using exclusively normal data or unlabeled data (mostly normal samples). One notorious issue with these approaches is that they are weak in discriminating anomalies from normal samples due to the lack of the knowledge about the anomalies. Here, we study the problem of few-shot anomaly detection, in which we aim at using a few labeled anomaly examples to train sample-efficient discriminative detection models. To address this problem, we introduce a novel weakly-supervised anomaly detection framework to train detection models without assuming the examples illustrating all possible classes of anomaly. Specifically, the proposed approach learns discriminative normality (regularity) by leveraging the labeled anomalies and a prior probability to enforce expressive representations of normality and unbounded deviated representations of abnormality. This is achieved by an end-to-end optimization of anomaly scores with a neural deviation learning, in which the anomaly scores of normal samples are imposed to approximate scalar scores drawn from the prior while that of anomaly examples is enforced to have statistically significant deviations from these sampled scores in the upper tail. Furthermore, our model is optimized to learn fine-grained normality and abnormality by top-K multiple-instance-learning-based feature subspace deviation learning, allowing more generalized representations. Comprehensive experiments on nine real-world image anomaly detection benchmarks show that our model is substantially more sample-efficient and robust, and performs significantly better than state-of-the-art competing methods in both closed-set and open-set settings. Our model can also offer explanation capability as a result of its prior-driven anomaly score learning. Code and datasets are available at: https://git.io/DevNet.
We propose an approach to instance segmentation from 3D point clouds based on dynamic convolution. This enables it to adapt, at inference, to varying feature and object scales. Doing so avoids some pitfalls of bottom up approaches, including a dependence on hyper-parameter tuning and heuristic post-processing pipelines to compensate for the inevitable variability in object sizes, even within a single scene. The representation capability of the network is greatly improved by gathering homogeneous points that have identical semantic categories and close votes for the geometric centroids. Instances are then decoded via several simple convolution layers, where the parameters are generated conditioned on the input. The proposed approach is proposal-free, and instead exploits a convolution process that adapts to the spatial and semantic characteristics of each instance. A light-weight transformer, built on the bottleneck layer, allows the model to capture long-range dependencies, with limited computational overhead. The result is a simple, efficient, and robust approach that yields strong performance on various datasets: ScanNetV2, S3DIS, and PartNet. The consistent improvements on both voxel- and point-based architectures imply the effectiveness of the proposed method. Code is available at: https://git.io/DyCo3D
Neural networks trained with SGD were recently shown to rely preferentially on linearly-predictive features and can ignore complex, equally-predictive ones. This simplicity bias can explain their lack of robustness out of distribution (OOD). The more complex the task to learn, the more likely it is that statistical artifacts (i.e. selection biases, spurious correlations) are simpler than the mechanisms to learn. We demonstrate that the simplicity bias can be mitigated and OOD generalization improved. We train a set of similar models to fit the data in different ways using a penalty on the alignment of their input gradients. We show theoretically and empirically that this induces the learning of more complex predictive patterns. OOD generalization fundamentally requires information beyond i.i.d. examples, such as multiple training environments, counterfactual examples, or other side information. Our approach shows that we can defer this requirement to an independent model selection stage. We obtain SOTA results in visual recognition on biased data and generalization across visual domains. The method - the first to evade the simplicity bias - highlights the need for a better understanding and control of inductive biases in deep learning.
Vision-and-Language Navigation (VLN) requires an agent to navigate to a remote location on the basis of natural-language instructions and a set of photo-realistic panoramas. Most existing methods take words in instructions and discrete views of each panorama as the minimal unit of encoding. However, this requires a model to match different textual landmarks in instructions (e.g., TV, table) against the same view feature. In this work, we propose an object-informed sequential BERT to encode visual perceptions and linguistic instructions at the same fine-grained level, namely objects and words, to facilitate the matching between visual and textual entities and hence "know what". Our sequential BERT enables the visual-textual clues to be interpreted in light of the temporal context, which is crucial to multi-round VLN tasks. Additionally, we enable the model to identify the relative direction (e.g., left/right/front/back) of each navigable location and the room type (e.g., bedroom, kitchen) of its current and final navigation goal, namely "know where", as such information is widely mentioned in instructions implying the desired next and final locations. Extensive experiments demonstrate the effectiveness compared against several state-of-the-art methods on three indoor VLN tasks: REVERIE, NDH, and R2R.
The limits of applicability of vision-and-language models are defined by the coverage of their training data. Tasks like vision question answering (VQA) often require commonsense and factual information beyond what can be learned from task-specific datasets. This paper investigates the injection of knowledge from general-purpose knowledge bases (KBs) into vision-and-language transformers. We use an auxiliary training objective that encourages the learned representations to align with graph embeddings of matching entities in a KB. We empirically study the relevance of various KBs to multiple tasks and benchmarks. The technique brings clear benefits to knowledge-demanding question answering tasks (OK-VQA, FVQA) by capturing semantic and relational knowledge absent from existing models. More surprisingly, the technique also benefits visual reasoning tasks (NLVR2, SNLI-VE). We perform probing experiments and show that the injection of additional knowledge regularizes the space of embeddings, which improves the representation of lexical and semantic similarities. The technique is model-agnostic and can expand the applicability of any vision-and-language transformer with minimal computational overhead.
Depression is among the most prevalent mental disorders, affecting millions of people of all ages globally. Machine learning techniques have shown effective in enabling automated detection and prediction of depression for early intervention and treatment. However, they are challenged by the relative scarcity of instances of depression in the data. In this work we introduce a novel deep multi-task recurrent neural network to tackle this challenge, in which depression classification is jointly optimized with two auxiliary tasks, namely one-class metric learning and anomaly ranking. The auxiliary tasks introduce an inductive bias that improves the classification model's generalizability on small depression samples. Further, unlike existing studies that focus on learning depression signs from static data without considering temporal dynamics, we focus on longitudinal data because i) temporal changes in personal development and family environment can provide critical cues for psychiatric disorders and ii) it may enable us to predict depression before the illness actually occurs. Extensive experimental results on child depression data show that our model is able to i) achieve nearly perfect performance in depression detection and ii) accurately predict depression 2-4 years before the clinical diagnosis, substantially outperforming seven competing methods.