Modern machine learning suffers from \textit{catastrophic forgetting} when learning new classes incrementally. The performance dramatically degrades due to the missing data of old classes. Incremental learning methods have been proposed to retain the knowledge acquired from the old classes, by using knowledge distilling and keeping a few exemplars from the old classes. However, these methods struggle to \textbf{scale up to a large number of classes}. We believe this is because of the combination of two factors: (a) the data imbalance between the old and new classes, and (b) the increasing number of visually similar classes. Distinguishing between an increasing number of visually similar classes is particularly challenging, when the training data is unbalanced. We propose a simple and effective method to address this data imbalance issue. We found that the last fully connected layer has a strong bias towards the new classes, and this bias can be corrected by a linear model. With two bias parameters, our method performs remarkably well on two large datasets: ImageNet (1000 classes) and MS-Celeb-1M (10000 classes), outperforming the state-of-the-art algorithms by 11.1\% and 13.2\% respectively.

In this paper, we propose an efficient and effective action recognition framework by combining multiple feature models from dynamic image, optical flow and raw frame, with 3D convolutional neural network (CNN). Dynamic image preserves the long-term temporal information, while optical flow captures short-term temporal information, and raw frame represents the appearance information. Experiments demonstrate that dynamic image provides complementary information to raw frame feature and optical flow feature. Furthermore, with the approximate rank pooling, the computation of dynamic images is about 360 times faster than optical flow, and the dynamic image requires far less memory than optical flow and raw frame.

Recently, with the advent of the convolutional neural network (CNN), many CNN-based object detection algorithms have been proposed and achieved encouraging results. In this paper, we introduce an algorithm based on region trajectories to establish the connections between object localizations in individual frames and video sequences. To detect object regions in the individual frames of a video, we enhance the region-based convolutional neural network (R-CNN), by incorporating EdgeBox with the Selective Search to generate candidate region proposals and combining the GoogLeNet with the AlexNet to improve the discriminability of the feature representations. The DeepMatching algorithm is employed in our proposed region trajectory method to track the points in the detected object regions. The experiments are conducted on the validation split of the TRECVID 2015 Localization dataset. As demonstrated by the experimental results, our proposed approach improves the object detection accuracy in both temporal and spatial measurements.

Recent studies have demonstrated that the implementation of local space-time interest points has good competence and robustness in the area of human action recognition, which has become one of the challenging problems in multimedia analysis. While most research focuses on the techniques of detecting feature points or capturing spatial and temporal information around those points, there has been very limited research on delving into the pooling strategies which are also important components of action recognition algorithms. In this paper, we propose a novel pooling framework by categorizing the interest points with respect to their idiosyncrasies. Specifically, we discuss three pooling strategies based on the optical flow orientation, foreground weight and spatio-temporal locations respectively and further investigate the fusion of different pooling strategies. For the encoding process, instead of the popular bag-of-visual words (BoV) method, we adopt the improved Fisher Vector (FV) approach. Our proposed methods are evaluated on a benchmark dataset with controlled settings (KTH), and two more challenging datasets with realistic background (HMDB51 and UCF101). The experimental results demonstrate that pooling strategies based on the appropriate idiosyncrasies of individual interest points can improve the performance of action classification.