Object detection using deep learning for real-time streaming applications
In various examples, techniques for optimizing object detection models are described herein. Systems and methods are disclosed that process sensor data using a backbone of a machine learning model(s) in order to generate feature maps at different resolutions. The systems and methods then use the machine learning model(s) to generate a vector based at least in part on one or more of the feature maps. For example, if the backbone generates four feature maps, then the machine learning model(s) may generate the vector using two feature maps from the four feature maps. The systems and methods then process the vector using a transformer of the machine learning model(s) in order to generate data representing a class label(s) for an object(s) depicted by an image represented by the sensor data and/or a location(s) of the object(s) within the image.
1 . A method comprising:
generating, using one or more first layers of a machine learning model and based at least on image data representing an image, a set of feature maps, each feature map of the set of feature maps corresponding to a different respective resolution of a set of resolutions;
determining a subset of the set of feature maps based at least on the subset of the set of feature maps being associated with one or more lowest resolutions of the set of resolutions;
generating, using one or more second layers of the machine learning model, a first vector to represent the subset of the set of feature maps;
generating, using an encoder of a transformer of the machine learning model and based at least on the first vector, a second vector; and
generating, using a decoder of the transformer of the machine learning model and based at least on the second vector, an output indicating at least one of a class label for an object depicted in the image or a location of the object within the image.
2 . The method of claim 1 , wherein the one or more first layers of the machine learning model comprise two or more first layers, the two or more first layers comprising:
at least a first layer of the two or more first layers to generate, based at least on the image data, a first feature map of the set of feature maps; and
at least a second layer of the two or more first layers to generate, based at least on the image data, a second feature map of the set of feature maps.
3 . The method of claim 1 , wherein the one or more first layers correspond to a ResNet feature extractor, and each feature map of the set of feature maps is generated using a separate instance of the ResNet feature extractor.
4 . The method of claim 1 , wherein:
the output is a third vector; and
the method further comprises generating, using a feed-forward network of the transformer of the machine learning model and based at least on the third vector, data representing at least one of the class label for the object depicted in the image or the location of the object within the image.
5 . The method of claim 1 , wherein:
the machine learning model uses the subset of the set of feature maps at the one or more lowest resolutions of the set of resolutions for object detection; and
sensor data for training the machine learning model is selected based at least on the machine learning model using the subset of the set of feature maps.
6 . One or more processors comprising processing circuitry to:
generate, using a ResNet feature extractor of a neural network, at least a first feature map at a first resolution using a first instance of the ResNet feature extractor and a second feature map at a second resolution using a second instance of the ResNet feature extractor;
generate, using the neural network and based at least on the first resolution and the second resolution, a vector representing the first feature map without representing the second feature map; and
compute, using a transformer of the neural network and based at least on the vector, one or more outputs indicative of at least one of classification information or location information corresponding to one or more objects.
7 . The one or more processors of claim 6 , wherein:
the neural network uses the first feature map at the first resolution without using the second feature map at the second resolution for object detection; and
training sensor data for training the neural network is selected based at least on the neural network using the first feature map without using the second feature map.
8 . The one or more processors of claim 7 , wherein:
the neural network further generates at least a third feature map at a third resolution; and
the first feature map and the third feature map are used to generate the vector based at least on one or more object types that the training sensor data corresponds to.
9 . The one or more processors of claim 6 , wherein the vector is further generated based at least on one or more object types that the neural network is configured to detect.
10 . The one or more processors of claim 6 , wherein:
the neural network further generates a third feature map at a third resolution and a fourth feature map at a fourth resolution; and
the vector is further generated to represent the third feature map without representing the fourth feature map.
11 . The one or more processors of claim 6 , wherein the neural network is trained using training data generated using a plurality of sensors that have at least one of varying poses, fields of view, or orientations.
12 . A system comprising:
one or more processors to:
generate, using one or more first layers of a machine learning model and based at least on image data representing an image, a set of feature maps, each feature map of the set of feature maps corresponding to a different resolution of a set of resolutions;
generate, using one or more second layers of the machine learning model and based at least on a subset of the set of feature maps being associated with one or more lowest resolutions of the set of resolution, a vector that represents the subset of the set of feature maps; and
generate, using a transformer of the machine learning model and based at least on the vector, an output indicating at least one of a class label for an object depicted in the image or a location of the object within the image.
13 . The system of claim 12 , wherein the one or more processors are to generate the set of feature maps using two or more first layers of the machine learning model, the two or more first layers comprising:
at least a first layer that generates, based at least on the image data, a first feature map of the set of feature maps at a first resolution of the set of resolutions; and
at least a second layer that generates, based at least on the image data, a second feature map of the set of features maps at a second resolution of the set of resolutions, the second resolution being different than the first resolution.
14 . The system of claim 12 , wherein the system is comprised in at least one of:
a control system for an autonomous or semi-autonomous machine;
a perception system for an autonomous or semi-autonomous machine;
a system for performing simulation operations;
a system for executing real-time streaming applications;
a system for generating at least one of virtual reality content, augmented reality content, or mixed reality content;
a system for performing digital twin operations;
a system for performing light transport simulation;
a system for performing collaborative content creation for 3D assets;
a system for performing deep learning operations;
a system implemented using an edge device;
a system implemented using a robot;
a system for performing conversational AI operations;
a system for generating synthetic data;
a system incorporating one or more virtual machines (VMs);
a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources.
15 . The method of claim 1 , wherein the determining the subset of the set of feature maps is further based at least on one or more object types being detected using the machine learning model.
16 . The method of claim 1 , wherein:
the set of feature maps includes four feature maps; and
the subset of the set of the feature maps includes one of two feature maps from the four feature maps or three feature maps from the four feature maps.
17 . The one or more processors of claim 6 , wherein the processing circuitry is further determine, based at least on one or more object types that the neural network is detecting, the first feature map for generating the vector.
18 . The system of claim 12 , wherein the vector is further generated based at least on one or more object types being detected using the machine learning model.
19 . The system of claim 12 , wherein:
the set of feature maps includes four feature maps; and
the subset of the set of the feature maps includes one of two feature maps from the four feature maps or three feature maps from the four feature maps.
20 . The system of claim 12 , wherein the one or more processors are further to at least one of:
select the subset of the set of feature maps to include one feature map associated with a lowest resolution of the set of resolutions; or
select the subset of the set of feature maps to include two feature maps associated with two lowest resolutions of the set of resolutions.