Salesforce, Inc.

Leland Stanford Junior University

Toyota Research Institute, Inc.

Juan Carlos Niebles Duque

Ran Xu

Caiming Xiong

Le Xue

Silvio Savarese

Chen Xing

Ning Yu

Adrien David Gaidon

Kuan-Hui Lee

Ehsan Adeli-Mosabbeb

Junnan Li

Shu Zhang

Nikhil Naik

Brian Chen

Karttikeya Mangalam

Ramprasaath Ramasamy Selvaraju

Roberto Martín-Martín

Boxiao Pan

Vibashan Vishnukumar Sharmini

Manli Shu

Haoye Cai

Honglu Zhou

De-An Huang

A method of training a neural network based three-dimensional (3D) encoder is provided. A first plurality of samples of a training dataset are generated using a first 3D model. An image generator with multi-view rendering is used to generate a plurality of two-dimensional (2D) images having different viewpoints of the first 3D model. A first language model is used to generate a plurality of texts corresponding to the plurality of 2D images respectively. A first text for a first image is generated by using one or more text descriptions generated by the first language model. A point cloud is generated by randomly sampling points in the 3D model. The first plurality of samples are generated using the plurality of 2D images, the corresponding plurality of texts, and the point cloud. The neural network based 3D encoder is trained using the training dataset including the first plurality of samples.

Systems and methods for multimodal pretraining for three-dimensional understanding models

A method of training a neural network based three-dimensional (3D) encoder is provided. A training dataset is generated using a plurality of 3D models of a 3D model dataset. To generate a first sample of the training dataset, an image generator with multi-view rendering is used to generate a plurality of image candidates of a first 3D model. A word is chosen from metadata associated with the first 3D model. A language model is used to generate one or more text descriptions using the selected word and a plurality of prompts. A point cloud is generated by randomly sampling points in the 3D model. The first sample is generated to include a first image randomly selected from the plurality of image candidates, one or more text descriptions, and the point cloud is generated. The 3D encoder is trained using the training dataset including the first sample.

Systems and methods for learning unified representations of language, image, and point cloud for three-dimensional recognition

Systems and methods for training a neural network based three-dimensional (3D) encoder for 3D classification are provided. A training dataset including a plurality of samples is received, wherein a first sample includes an image, a text, and a point cloud. An image encoder of a pretrained vision and language model is used to generate image representations for the image of the first sample. A text encoder of the pretrained vision and language model is used to generate text representations for the text of the first sample. The neural network based 3D encoder is used to generate 3D representations for the point cloud of the first sample. A loss objective is computed based on the image representations, text representations, and 3D representations. Parameters of the neural network based 3D encoder are updated based on the computed loss objective via backpropagation.

Embodiments described herein provide a system for three-dimensional (3D) object detection. The system includes an input interface configured to obtain 3D point data describing spatial information of a plurality of points, and a memory storing a neural network based 3D object detection model having an encoder and a decoder. The system also includes processors to perform operations including: encoding, by the encoder, a first set of coordinates into a first set of point features and a set of object features; sampling a second set of point features from the first set of point features; generating, by attention layers at the decoder, a set of attention weights by applying cross-attention over at least the set of object features and the second set of point feature, and generate, by the decoder, a predicted bounding box among the plurality of points based on at least in part on the set of attention weights.

Systems and methods for attention mechanism in three-dimensional object detection

Embodiments described herein provide an open-vocabulary instance segmentation framework that adopts a pre-trained vision-language model to develop a pipeline in detecting novel categories of instances.

Systems and methods for open vocabulary instance segmentation in unannotated images

Embodiments described herein provide systems and methods for training video models to perform a task from an input instructional video. A procedure knowledge graph (PKG) may be generated with nodes representing procedure steps, and edges representing relationships between the steps. The PKG may be generated based on text and/or video training data which includes procedures (e.g., instructional videos). Using the PKG, a video model may be trained using the PKG to provide supervisory training signals for a number of tasks. Once the model is trained, it may be fine-tuned for a specific task which benefits from the model being trained in a way that makes the model embed procedural information when encoding videos.

Systems and methods for video models with procedure understanding

Embodiments described herein provide an intelligent method to select instances, by utilizing unsupervised tracking for videos. Using this freely available form of supervision, a temporal constraint is adopted for selecting instances that ensures that different instances contain the same object while sampling the temporal augmentation from the video. In addition, using the information on the spatial extent of the tracked object, spatial constraints are applied to ensure that sampled instances overlap meaningfully with the tracked object. Taken together, these spatiotemporal constraints result in better supervisory signal for contrastive learning from videos.

Systems and methods for contrastive pretraining with video tracking supervision

A method for scene perception using video captioning based on a spatio-temporal graph model is described. The method includes decomposing the spatio-temporal graph model of a scene in input video into a spatial graph and a temporal graph. The method also includes modeling a two branch framework having an object branch and a scene branch according to the spatial graph and the temporal graph to learn object interactions between the object branch and the scene branch. The method further includes transferring the learned object interactions from the object branch to the scene branch as privileged information. The method also includes captioning the scene by aligning language logits from the object branch and the scene branch according to the learned object interactions.

Spatio-temporal graph for video captioning with knowledge distillation

A method for predicting spatial positions of several key points on a human body in the near future in an egocentric setting is described. The method includes generating a frame-level supervision for human poses. The method also includes suppressing noise and filling missing joints of the human body using a pose completion module. The method further includes splitting the poses into a global stream and a local stream. Furthermore, the method includes combining the global stream and the local stream to forecast future human locomotion.

Growing community of inventors

Mountain View, CA, United States of America

Juan Carlos Niebles Duque