This paper reveals a characteristic of DEtection Transformer (DETR) that negatively impacts its training efficacy, i.e., the cross-attention and self-attention layers in DETR decoder have contrary impacts on the object queries (though both impacts are important). Specifically, we observe the cross-attention tends to gather multiple queries around the same object, while the self-attention disperses these queries far away. To improve the training efficacy, we propose a Divide-And-Conquer DETR (DAC-DETR) that divides the cross-attention out from this contrary for better conquering. During training, DAC-DETR employs an auxiliary decoder that focuses on learning the cross-attention layers. The auxiliary decoder, while sharing all the other parameters, has NO self-attention layers and employs one-to-many label assignment to improve the gathering effect. Experiments show that DAC-DETR brings remarkable improvement over popular DETRs. For example, under the 12 epochs training scheme on MS-COCO, DAC-DETR improves Deformable DETR (ResNet-50) by +3.4 AP and achieves 50.9 (ResNet-50) / 58.1 AP (Swin-Large) based on some popular methods (i.e., DINO and an IoU-related loss).

This paper reveals a characteristic of DEtection Transformer (DETR) that negatively impacts its training efficacy, i.e., the cross-attention and self-attention layers in DETR decoder have contrary impacts on the object queries (though both impacts are important). Specifically, we observe the cross-attention tends to gather multiple queries around the same object, while the self-attention disperses these queries far away. To improve the training efficacy, we propose a Divide-And-Conquer DETR (DAC-DETR) that divides the cross-attention out from this contrary for better conquering. During training, DAC-DETR employs an auxiliary decoder that focuses on learning the cross-attention layers. The auxiliary decoder, while sharing all the other parameters, has NO self-attention layers and employs one-to-many label assignment to improve the gathering effect.

As a cost-effective and robust technology, automotive radar has seen steady improvement during the last years, making it an appealing complement to commonly used sensors like camera and LiDAR in autonomous driving. Radio frequency data with rich semantic information are attracting more and more attention. Most current radar-based models take radio frequency image sequences as the input. However, these models heavily rely on convolutional neural networks and leave out the spatial-temporal semantic context during the encoding stage. To solve these problems, we propose a model called Mask-RadarNet to fully utilize the hierarchical semantic features from the input radar data. Mask-RadarNet exploits the combination of interleaved convolution and attention operations to replace the traditional architecture in transformer-based models. In addition, patch shift is introduced to the Mask-RadarNet for efficient spatial-temporal feature learning. By shifting part of patches with a specific mosaic pattern in the temporal dimension, Mask-RadarNet achieves competitive performance while reducing the computational burden of the spatial-temporal modeling. In order to capture the spatial-temporal semantic contextual information, we design the class masking attention module (CMAM) in our encoder. Moreover, a lightweight auxiliary decoder is added to our model to aggregate prior maps generated from the CMAM. Experiments on the CRUW dataset demonstrate the superiority of the proposed method to some state-of-the-art radar-based object detection algorithms. With relatively lower computational complexity and fewer parameters, the proposed Mask-RadarNet achieves higher recognition accuracy for object detection in autonomous driving.

In this work, we propose a novel deformable convolutional pyramid network for unsupervised image registration. Specifically, the proposed network enhances the traditional pyramid network by adding an additional shared auxiliary decoder for image pairs. This decoder provides multi-scale high-level feature information from unblended image pairs for the registration task. During the registration process, we also design a multi-scale feature fusion block to extract the most beneficial features for the registration task from both global and local contexts. Validation results indicate that this method can capture complex deformations while achieving higher registration accuracy and maintaining smooth and plausible deformations.

Although robot-assisted cardiovascular catheterization is commonly performed for intervention of cardiovascular diseases, more studies are needed to support the procedure with automated tool segmentation. This can aid surgeons on tool tracking and visualization during intervention. Learning-based segmentation has recently offered state-of-the-art segmentation performances however, generating ground-truth signals for fully-supervised methods is labor-intensive and time consuming for the interventionists. In this study, a weakly-supervised learning method with multi-lateral pseudo labeling is proposed for tool segmentation in cardiac angiograms. The method includes a modified U-Net model with one encoder and multiple lateral-branched decoders that produce pseudo labels as supervision signals under different perturbation. The pseudo labels are self-generated through a mixed loss function and shared consistency in the decoders. We trained the model end-to-end with weakly-annotated data obtained during robotic cardiac catheterization. Experiments with the proposed model shows weakly annotated data has closer performance to when fully annotated data is used. Compared to three existing weakly-supervised methods, our approach yielded higher segmentation performance across three different cardiac angiogram data. With ablation study, we showed consistent performance under different parameters. Thus, we offer a less expensive method for real-time tool segmentation and tracking during robot-assisted cardiac catheterization.

We investigate whisper-to-natural-speech conversion using sequence-to-sequence approach by proposing modified transformer architecture. We investigate different features like mel frequency cepstral coefficients and smoothed spectral features. The proposed networks are trained end-to-end using supervised approach for feature-to-feature transformation. Further, We also investigate the effectiveness of embedded auxiliary decoder used after N encoder sub-layers, and is trained with the frame-level objective function for identifying source phoneme labels. We show results on wTIMIT and CHAINS datasets by measuring word error rate using end-to-end ASR and also BLEU scores for the generated speech. In addition, we measure spectral shape of it by measuring formant distributions w.r.t. the reference speech, as formant divergence metric. We have found whisper-to-natural converted speech formants probability distribution is similar to the ground-truth distribution. To the authors' best knowledge, this is the first time modified transformer has been applied for whisper-to-natural-speech conversion and vice versa.

Variational autoencoders learn unsupervised data representations, but these models frequently converge to minima that fail to preserve meaningful semantic information. For example, variational autoencoders with autoregressive decoders often collapse into autodecoders, where they learn to ignore the encoder input. In this work, we demonstrate that adding an auxiliary decoder to regularize the latent space can prevent this collapse, but successful auxiliary decoding tasks are domain dependent. Auxiliary decoders can increase the amount of semantic information encoded in the latent space and visible in the reconstructions. The semantic information in the variational autoencoder's representation is only weakly correlated with its rate, distortion, or evidence lower bound. Compared to other popular strategies that modify the training objective, our regularization of the latent space generally increased the semantic information content.

Autoregressive generative models of images tend to be biased towards capturing local structure, and as a result they often produce samples which are lacking in terms of large-scale coherence. To address this, we propose two methods to learn discrete representations of images which abstract away local detail. We show that autoregressive models conditioned on these representations can produce high-fidelity reconstructions of images, and that we can train autoregressive priors on these representations that produce samples with large-scale coherence. We can recursively apply the learning procedure, yielding a hierarchy of progressively more abstract image representations. We train hierarchical class-conditional autoregressive models on the ImageNet dataset and demonstrate that they are able to generate realistic images at resolutions of 128$\times$128 and 256$\times$256 pixels.