In video prediction tasks, one major challenge is to capture the multi-modal nature of future contents and dynamics. In this work, we propose a simple yet effective framework that can efficiently predict plausible future states. The key insight is that the potential distribution of a sequence could be approximated with analogous ones in a repertoire of training pool, namely, expert examples. By further incorporating a novel optimization scheme into the training procedure, plausible predictions can be sampled efficiently from distribution constructed from the retrieved examples. Meanwhile, our method could be seamlessly integrated with existing stochastic predictive models; significant enhancement is observed with comprehensive experiments in both quantitative and qualitative aspects. We also demonstrate the generalization ability to predict the motion of unseen class, i.e., without access to corresponding data during training phase.
Videos of actions are complex spatio-temporal signals, containing rich compositional structures. Current generative models are limited in their ability to generate examples of object configurations outside the range they were trained on. Towards this end, we introduce a generative model (AG2Vid) based on Action Graphs, a natural and convenient structure that represents the dynamics of actions between objects over time. Our AG2Vid model disentangles appearance and position features, allowing for more accurate generation. AG2Vid is evaluated on the CATER and Something-Something datasets and outperforms other baselines. Finally, we show how Action Graphs can be used for generating novel compositions of unseen actions.
Faced with new and different data during testing, a model must adapt itself. We consider the setting of fully test-time adaptation, in which a supervised model confronts unlabeled test data from a different distribution, without the help of its labeled training data. We propose an entropy minimization approach for adaptation: we take the model's confidence as our objective as measured by the entropy of its predictions. During testing, we adapt the model by modulating its representation with affine transformations to minimize entropy. Our experiments show improved robustness to corruptions for image classification on CIFAR-10/100 and ILSVRC and demonstrate the feasibility of target-only domain adaptation for digit classification on MNIST and SVHN.
Similarity metrics for instances have drawn much attention, due to their importance for computer vision problems such as object tracking. However, existing methods regard object similarity learning as a post-hoc stage after object detection and only use sparse ground truth matching as the training objective. This process ignores the majority of the regions on the images. In this paper, we present a simple yet effective quasi-dense matching method to learn instance similarity from hundreds of region proposals in a pair of images. In the resulting feature space, a simple nearest neighbor search can distinguish different instances without bells and whistles. When applied to joint object detection and tracking, our method can outperform existing methods without using location or motion heuristics, yielding almost 10 points higher MOTA on BDD100K and Waymo tracking datasets. Our method is also competitive on one-shot object detection, which further shows the effectiveness of quasi-dense matching for category-level metric learning. The code will be available at https://github.com/sysmm/quasi-dense.
Metric learning seeks perceptual embeddings where visually similar instances are close and dissimilar instances are apart, but learn representation can be sub-optimal when the distribution of intra-class samples is diverse and distinct sub-clusters are present. We theoretically prove and empirically show that under reasonable noise assumptions, prevalent embedding losses in metric learning, e.g., triplet loss, tend to project all samples of a class with various modes onto a single point in the embedding space, resulting in class collapse that usually renders the space ill-sorted for classification or retrieval. To address this problem, we propose a simple modification to the embedding losses such that each sample selects its nearest same-class counterpart in a batch as the positive element in the tuple. This allows for the presence of multiple sub-clusters within each class. The adaptation can be integrated into a wide range of metric learning losses. Our method demonstrates clear benefits on various fine-grained image retrieval datasets over a variety of existing losses; qualitative retrieval results show that samples with similar visual patterns are indeed closer in the embedding space.
We investigate the problem of predicting driver behavior in parking lots, an environment which is less structured than typical road networks and features complex, interactive maneuvers in a compact space. Using the CARLA simulator, we develop a parking lot environment and collect a dataset of human parking maneuvers. We then study the impact of model complexity and feature information by comparing a multi-modal Long Short-Term Memory (LSTM) prediction model and a Convolution Neural Network LSTM (CNN-LSTM) to a physics-based Extended Kalman Filter (EKF) baseline. Our results show that 1) intent can be estimated well (roughly 85% top-1 accuracy and nearly 100% top-3 accuracy with the LSTM and CNN-LSTM model); 2) knowledge of the human driver's intended parking spot has a major impact on predicting parking trajectory; and 3) the semantic representation of the environment improves long term predictions.
Computer vision algorithms, e.g. for face recognition, favour groups of individuals that are better represented in the training data. This happens because of the generalization that classifiers have to make. It is simpler to fit the majority groups as this fit is more important to overall error. We propose to create a balanced training dataset, consisting of the original dataset plus new data points in which the group memberships are intervened, minorities become majorities and vice versa. We show that current generative adversarial networks are a powerful tool for learning these data points, called contrastive examples. We experiment with the equalized odds bias measure on tabular data as well as image data (CelebA and Diversity in Faces datasets). Contrastive examples allow us to expose correlations between group membership and other seemingly neutral features. Whenever a causal graph is available, we can put those contrastive examples in the perspective of counterfactuals.
Meta-learning has become a popular framework for few-shot learning in recent years, with the goal of learning a model from collections of few-shot classification tasks. While more and more novel meta-learning models are being proposed, our research has uncovered simple baselines that have been overlooked. We present a Meta-Baseline method, by pre-training a classifier on all base classes and meta-learning on a nearest-centroid based few-shot classification algorithm, it outperforms recent state-of-the-art methods by a large margin. Why does this simple method work so well? In the meta-learning stage, we observe that a model generalizing better on unseen tasks from base classes can have a decreasing performance on tasks from novel classes, indicating a potential objective discrepancy. We find both pre-training and inheriting a good few-shot classification metric from the pre-trained classifier are important for Meta-Baseline, which potentially helps the model better utilize the pre-trained representations with stronger transferability. Furthermore, we investigate when we need meta-learning in this Meta-Baseline. Our work sets up a new solid benchmark for this field and sheds light on further understanding the phenomenons in the meta-learning framework for few-shot learning.
Spatio-temporal action detection in videos requires localizing the action both spatially and temporally in the form of an "action tube". Nowadays, most spatio-temporal action detection datasets (e.g. UCF101-24, AVA, DALY) are annotated with action tubes that contain a single person performing the action, thus the predominant action detection models simply employ a person detection and tracking pipeline for localization. However, when the action is defined as an interaction between multiple objects, such methods may fail since each bounding box in the action tube contains multiple objects instead of one person. In this paper, we study the spatio-temporal action detection problem with multi-object interaction. We introduce a new dataset that is annotated with action tubes containing multi-object interactions. Moreover, we propose an end-to-end spatio-temporal action detection model that performs both spatial and temporal regression simultaneously. Our spatial regression may enclose multiple objects participating in the action. During test time, we simply connect the regressed bounding boxes within the predicted temporal duration using a simple heuristic. We report the baseline results of our proposed model on this new dataset, and also show competitive results on the standard benchmark UCF101-24 using only RGB input.
Weakly-supervised action localization problem requires training a model to localize the action segments in the video given only video level action label. It can be solved under the Multiple Instance Learning (MIL) framework, where a bag (video) contains multiple instances (action segments). Since only the bag's label is known, the main challenge is to assign which key instances within the bag trigger the bag's label. Most previous models use an attention-based approach. These models use attention to generate the bag's representation from instances and then train it via bag's classification. In this work, we explicitly model the key instances assignment as a hidden variable and adopt an Expectation-Maximization framework. We derive two pseudo-label generation schemes to model the E and M process and iteratively optimize the likelihood lower bound. We also show that previous attention-based models implicitly violate the MIL assumptions that instances in negative bags should be uniformly negative. In comparison, Our EM-MIL approach more accurately models these assumptions. Our model achieves state-of-the-art performance on two standard benchmarks, THUMOS14 and ActivityNet1.2, and shows the superiority of detecting relative complete action boundary in videos containing multiple actions.