Exploring refined dual visual features cross-combination for image captioning.

For current image caption tasks used to encode region features and grid features Transformer-based encoders have become commonplace, because of their multi-head self-attention mechanism, the encoder can better capture the relationship between different regions in the image and contextual information. However, stacking Transformer blocks necessitates quadratic computation through self-attention to visual features, not only resulting in the computation of numerous redundant features but also significantly increasing computational overhead. This paper presents a novel Distilled Cross-Combination Transformer (DCCT) network. Technically, we first introduce a distillation cascade fusion encoder (DCFE), where a probabilistic sparse self-attention layer is used to filter out some redundant and distracting features that affect attention focus, aiming to obtain more refined visual features and enhance encoding efficiency. Next, we develop a parallel cross-fusion attention module (PCFA) that fully exploits the complementarity and correlation between grid and region features to better fuse the encoded dual visual features. Extensive experiments conducted on the MSCOCO dataset demonstrate that our proposed DCCT method achieves outstanding performance, rivaling current state-of-the-art approaches.

SiamBAN: Target-Aware Tracking With Siamese Box Adaptive Network.

Variation of scales or aspect ratios has been one of the main challenges for tracking. To overcome this challenge, most existing methods adopt either multi-scale search or anchor-based schemes, which use a predefined search space in a handcrafted way and therefore limit their performance in complicated scenes. To address this problem, recent anchor-free based trackers have been proposed without using prior scale or anchor information. However, an inconsistency problem between classification and regression degrades the tracking performance. To address the above issues, we propose a simple yet effective tracker (named Siamese Box Adaptive Network, SiamBAN) to learn a target-aware scale handling schema in a data-driven manner. Our basic idea is to predict the target boxes in a per-pixel fashion through a fully convolutional network, which is anchor-free. Specifically, SiamBAN divides the tracking problem into classification and regression tasks, which directly predict objectiveness and regress bounding boxes, respectively. A no-prior box design is proposed to avoid tuning hyper-parameters related to candidate boxes, which makes SiamBAN more flexible. SiamBAN further uses a target-aware branch to address the inconsistency problem. Experiments on benchmarks including VOT2018, VOT2019, OTB100, UAV123, LaSOT and TrackingNet show that SiamBAN achieves promising performance and runs at 35 FPS.

Improving Alzheimer's Disease Classification by Combining Multiple Measures.

Several anatomical magnetic resonance imaging (MRI) markers for Alzheimer's disease (AD) have been identified. Cortical gray matter volume, cortical thickness, and subcortical volume have been used successfully to assist the diagnosis of Alzheimer's disease including its early warning and developing stages, e.g., mild cognitive impairment (MCI) including MCI converted to AD (MCIc) and MCI not converted to AD (MCInc). Currently, these anatomical MRI measures have mainly been used separately. Thus, the full potential of anatomical MRI scans for AD diagnosis might not yet have been used optimally. Meanwhile, most studies currently only focused on morphological features of regions of interest (ROIs) or interregional features without considering the combination of them. To further improve the diagnosis of AD, we propose a novel approach of extracting ROI features and interregional features based on multiple measures from MRI images to distinguish AD, MCI (including MCIc and MCInc), and health control (HC). First, we construct six individual networks based on six different anatomical measures (i.e., CGMV, CT, CSA, CC, CFI, and SV) and Automated Anatomical Labeling (AAL) atlas for each subject. Then, for each individual network, we extract all node (ROI) features and edge (interregional) features, and denoted as node feature set and edge feature set, respectively. Therefore, we can obtain six node feature sets and six edge feature sets from six different anatomical measures. Next, each feature within a feature set is ranked by -score in descending order, and the top ranked features of each feature set are applied to MKBoost algorithm to obtain the best classification accuracy. After obtaining the best classification accuracy, we can get the optimal feature subset and the corresponding classifier for each node or edge feature set. Afterwards, to investigate the classification performance with only node features, we proposed a weighted multiple kernel learning (wMKL) framework to combine these six optimal node feature subsets, and obtain a combined classifier to perform AD classification. Similarly, we can obtain the classification performance with only edge features. Finally, we combine both six optimal node feature subsets and six optimal edge feature subsets to further improve the classification performance. Experimental results show that the proposed method outperforms some state-of-the-art methods in AD classification, and demonstrate that different measures contain complementary information.