Chao-Yuan Wu’s research while affiliated with Meta and other places

... Instead of directly applying it for diffusion learning , our novel design concretizes the unordered tokens into the shape of the 3D input. Specifically, this is achieved by cross-attending (Huang et al., 2024b) the set latent via a sparse point cloud sampled from the input 3D shape, as visualized in Fig. 2. The resulting point-cloud structured latent space significantly facilitate shape-texture disentanglement and 3D editing. Afterward, a DiT-based 3D decoder (Peebles & Xie, 2023; gradually decodes and upsamples the latent point cloud into a set of dense surfel Gaussians (Huang et al., 2024a), which are rasterized to high-resolution renderings to supervise 3D VAE training. ...

Reference:

GaussianAnything: Interactive Point Cloud Latent Diffusion for 3D Generation

... Both deep learning architectures are extensively employed in the medical image segmentation area and have realized satisfactory results. In addition to classic networks, recent models like Segment Anything Model (SAM) [18] and SAM2 [19] which trained with billions of annotated images have also demonstrated generalizable performances in image segmentation tasks. ...

Reference:

Robust Real-time Segmentation of Bio-Morphological Features in Human Cherenkov Imaging during Radiotherapy via Deep Learning

... Earlier works [10,19,25,37,59,62,66] in this field typically predict only the geometry and train on small datasets [5,50], which limits their generalization ability. Recently, larger 3D datasets [11,42] have been collected, unlocking the potential to train feedforward 3D models at scale [24,27,61]. These models exhibit great generalization ability to unseen images, and excel at producing reconstructions that tightly align with the observed cues in the input image. ...

Reference:

SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images

... We used off-the-shelf MViT2-Huge [41] in the Detectron2 framework [42] trained on MS-COCO [43] as the 2D object detector and Metric3DV2giant [9] for metric depth estimation. We use our pseudo-labels to train MonoDETR [19] monocular 3D object detector, using AdamW [44] optimiser with learning rate and weight decay equal to 0.0002 and 0.0001, respectively, while keeping other hyper-parameters as in [19]. ...

Reference:

MonoSOWA: Scalable monocular 3D Object detector Without human Annotations

... Figure 4 presents saliency maps highlighting regions influencing the stenoses predictions. Additional performance metrics using MemVIT architecture [30] instead of the R2D + 1 are presented in Supplemental Materials Table S13. ...

Reference:

Automated stenosis estimation of coronary angiographies using end-to-end learning

... Deep learning, and particularly convolutional neural networks (CNNs), has gained traction in radar localization due to their strong performance in extracting detailed features [17,18]. While CNNs can capture informative features for estimation, their focus on localized relationships limits their ability to fully capture global dependencies across arrays, which is critical in FDA-based localization [19]. ...

Reference:

Globally fused hierarchical transformer for nonuniform frequency diverse arrays

... Some approaches [11,13,42] extend MAE [12] to the video field. MaskFeat [43] studies five different types of features in reconstructed masked regions. Hiera [15] combines MAE pre-training with a hierarchical transformer architecture in a simple way. ...

Reference:

ST-HViT: spatial-temporal hierarchical vision transformer for action recognition

... Several techniques have been developed to optimize GPU memory utilization during LLM training. Reversible subnetworks (Liao, Tan, and Monz 2023;Mangalam et al. 2022;Zhao et al. 2024a;Gomez et al. 2017b;Kitaev, Kaiser, and Levskaya 2020) minimize activation memory by recalculating activations during backpropagation. Gradient checkpointing (Chen et al. 2016) improves memory efficiency by discarding and later reconstructing some intermediate activations through an additional forward pass. ...

Reference:

Gradient Weight-normalized Low-rank Projection for Efficient LLM Training

... 8,9 Convolutional Neural Networks (CNNs), as introduced by LeCun et al. in 1989 10 and popularized by Krizhevsky et al. in 2012, have brought transformative advancements to medical imaging. 10,11 These networks have revolutionized the field by harnessing their capacity to autonomously learn intricate data representations. The renaissance of CNNs has yielded remarkable progress across diverse medical imaging modalities, including Radiography, 12 Endoscopy, 13 Computed Tomography(CT). ...

Reference:

Deep learning models for tumor detection and segmentation in medical image analysis: A comprehensive review of ResNet, U-Net, DETR, and inception variants

... This method applies a certain percentage of the mask to images, then uses an encoder to learn the visible patch features, and a decoder to reconstruct the original image, continuously optimizing the model's ability to learn general features. The reconstruction target has many forms, such as pixels [10,11], discretized tokens [12,13], gradient histograms [14], features generated by other networks [15][16][17][18] and features extracted by a teacher network [19]. Currently, the method with better performance and efficiency in the natural image field is the last one. ...

Reference:

PCB CT Image Element Segmentation Model Optimizing the Semantic Perception of Connectivity Relationship