3D CNN shows its strong ability in learning spatiotemporal representation in recent video recognition tasks. However, inflating 2D convolution to 3D inevitably introduces additional computational costs, making it cumbersome in practical deployment. We consider whether there is a way to equip the conventional 2D convolution with temporal vision no requiring expanding its kernel. To this end, we propose the video shuffle, a parameter-free plug-in component that efficiently reallocates the inputs of 2D convolution so that its receptive field can be extended to the temporal dimension. In practical, video shuffle firstly divides each frame feature into multiple groups and then aggregate the grouped features via temporal shuffle operation. This allows the following 2D convolution aggregate the global spatiotemporal features. The proposed video shuffle can be flexibly inserted into popular 2D CNNs, forming the Video Shuffle Networks (VSN). With a simple yet efficient implementation, VSN performs surprisingly well on temporal modeling benchmarks. In experiments, VSN not only gains non-trivial improvements on Kinetics and Moments in Time, but also achieves state-of-the-art performance on Something-Something-V1, Something-Something-V2 datasets.
We present DeepICP - a novel end-to-end learning-based 3D point cloud registration framework that achieves comparable registration accuracy to prior state-of-the-art geometric methods. Different from other keypoint based methods where a RANSAC procedure is usually needed, we implement the use of various deep neural network structures to establish an end-to-end trainable network. Our keypoint detector is trained through this end-to-end structure and enables the system to avoid the inference of dynamic objects, leverages the help of sufficiently salient features on stationary objects, and as a result, achieves high robustness. Rather than searching the corresponding points among existing points, the key contribution is that we innovatively generate them based on learned matching probabilities among a group of candidates, which can boost the registration accuracy. Our loss function incorporates both the local similarity and the global geometric constraints to ensure all above network designs can converge towards the right direction. We comprehensively validate the effectiveness of our approach using both the KITTI dataset and the Apollo-SouthBay dataset. Results demonstrate that our method achieves comparable or better performance than the state-of-the-art geometry-based methods. Detailed ablation and visualization analysis are included to further illustrate the behavior and insights of our network. The low registration error and high robustness of our method makes it attractive for substantial applications relying on the point cloud registration task.
The unprecedented demand for large amount of data has catalyzed the trend of combining human insights with machine learning techniques, which facilitate the use of crowdsourcing to enlist label information both effectively and efficiently. The classic work on crowdsourcing mainly focuses on the label inference problem under the categorization setting. However, inferring the true label requires sophisticated aggregation models that usually can only perform well under certain assumptions. Meanwhile, no matter how complicated the aggregation model is, the true model that generated the crowd labels remains unknown. Therefore, the label inference problem can never infer the ground truth perfectly. Based on the fact that the crowdsourcing labels are abundant and utilizing aggregation will lose such kind of rich annotation information (e.g., which worker provided which labels), we believe that it is critical to take the diverse labeling abilities of the crowdsourcing workers as well as their correlations into consideration. To address the above challenge, we propose to tackle three research problems, namely inference, learning, and teaching.
With the increasing demand for large amount of labeled data, crowdsourcing has been used in many large-scale data mining applications. However, most existing works in crowdsourcing mainly focus on label inference and incentive design. In this paper, we address a different problem of adaptive crowd teaching, which is a sub-area of machine teaching in the context of crowdsourcing. Compared with machines, human beings are extremely good at learning a specific target concept (e.g., classifying the images into given categories) and they can also easily transfer the learned concepts into similar learning tasks. Therefore, a more effective way of utilizing crowdsourcing is by supervising the crowd to label in the form of teaching. In order to perform the teaching and expertise estimation simultaneously, we propose an adaptive teaching framework named JEDI to construct the personalized optimal teaching set for the crowdsourcing workers. In JEDI teaching, the teacher assumes that each learner has an exponentially decayed memory. Furthermore, it ensures comprehensiveness in the learning process by carefully balancing teaching diversity and learner's accurate learning in terms of teaching usefulness. Finally, we validate the effectiveness and efficacy of JEDI teaching in comparison with the state-of-the-art techniques on multiple data sets with both synthetic learners and real crowdsourcing workers.
We describe an attentive encoder that combines tree-structured recursive neural networks and sequential recurrent neural networks for modelling sentence pairs. Since existing attentive models exert attention on the sequential structure, we propose a way to incorporate attention into the tree topology. Specially, given a pair of sentences, our attentive encoder uses the representation of one sentence, which generated via an RNN, to guide the structural encoding of the other sentence on the dependency parse tree. We evaluate the proposed attentive encoder on three tasks: semantic similarity, paraphrase identification and true-false question selection. Experimental results show that our encoder outperforms all baselines and achieves state-of-the-art results on two tasks.