Training a machine-learned model to detect low variance regions
Low variance detection training is described herein. In an example, annotated data can be determined based on sensor data received from a sensor associated with a vehicle. The annotated data can comprise an annotated low variance region and/or an annotated high variance region. The sensor data can be input into a model, and the model can determine an output comprising a high variance output and a low variance output. In an example, a difference between the annotated data and the output can be determined and one or more parameters associated with the model can be altered based at least in part on the difference. The model can be transmitted to a vehicle configured to be controlled by another output of the model.
1. A method comprising:
receiving data associated with environments of vehicles;
determining annotated data based at least in part on the data, wherein the annotated data comprises an annotated high variance region in the data and an annotated low variance region in the data;
inputting the data into a model;
determining, by the model, an output comprising a low variance output including a first feature detection and a high variance output including a second feature detection based on the first feature detection, the model trained to determine the low variance output;
determining a difference between the output and the annotated data;
altering one or more parameters of the model based at least in part on the difference; and
transmitting the model to a vehicle configured to be controlled by another output of the model.
2. The method as claim 1 recites, wherein the annotated low variance region is determined from one or more statistical models.
3. The method as claim 1 recites, wherein the annotated low variance region is determined based at least in part on a feature associated with the data.
4. The method as claim 3 recites, the method further comprising:
inputting the feature into an additional model;
receiving, from the additional model, a reconstructed output; and
determining a loss based on a difference between the reconstructed output and the data,
wherein altering the one or more parameters is further based at least in part on the loss.
5. The method as claim 1 recites, wherein:
the model is a neural network, and
the high variance output is based on the low variance output,
the method further comprising determining an additional high variance output, and
further wherein altering the one or more parameters comprises training the model end-to-end based at least in part on the low variance output, the high variance output, and the additional high variance output.
6. The method as claim 1 recites, wherein the first feature detection is a head detection and the second feature detection is a pedestrian detection.
7. The method as claim 1 recites, wherein the data comprises image data, a batch of image data, or an image space.
8. A system comprising:
one or more processors; and
one or more non-transitory computer-readable media that, when executed by the one or more processors, cause the system to perform operations comprising:
receiving sensor data from a sensor associated with environments of vehicles;
annotating, as annotated data and based on the sensor data, an annotated low variance region associated with the sensor data;
training a model based at least in part on the annotated data and the sensor data to generate a trained model, the trained model configured to output an indication of a low variance region including a first feature detection and an indication of a high variance region including a second feature detection based at least in part on the first feature detection; and
transmitting the trained model to a vehicle configured to be controlled by another output of the model.
9. The system as claim 8 recites, wherein the sensor data comprises at least one of image data, a batch of image data, or an image space.
10. The system as claim 9 recites, wherein the annotated low variance region is associated with the first feature detection of the sensor data determined by the model.
11. The system as claim 10 recites, the operations further comprising:
mapping the first feature detection to a reconstructed input; and
determining, as a loss, a difference between the sensor data and the reconstructed input,
wherein training the model is further based at least in part on the loss.
12. The system as claim 11 recites, wherein mapping the first feature detection to the reconstructed input comprises:
inputting the first feature detection into an additional model; and
receiving, from the additional model, the reconstructed input.
13. The system as claim 11 recites,
wherein training the model comprises training the model from end to end.
14. The system as claim 8 recites, wherein the annotated low variance region is determined from one or more statistical models based at least in part on one or more of entropy, pixel intensity, or aspect ratios associated with indications of low variance regions.
15. One or more non-transitory computer-readable media that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
determining, based on sensor data received from a sensor associated with a vehicle, annotated data, wherein the annotated data comprises one or more of an annotated low variance region or an annotated high variance region;
inputting the sensor data into a model;
determining, by the model, an output comprising a low variance output including a first feature detection and a high variance output including a second feature detection based on the first feature detection, wherein the low variance output indicates a region of sensor data associated with at least a portion of an object that varies from object to object, within a same object classification, at or below a threshold variance;
determining a difference between the annotated data and the output;
altering one or more parameters associated with the model based at least in part on the difference; and
transmitting the model to a vehicle configured to be controlled by another output of the model.
16. The one or more non-transitory computer-readable media as claim 15 recites, wherein determining the annotated data comprises determining the annotated low variance region based at least in part on a statistical model associated with one or more of the sensor data or an intermediary output of the model based at least in part on the sensor data.
17. The one or more non-transitory computer-readable media as claim 15 recites, the operations further comprising:
inputting at least a portion of the sensor data into the model;
receiving, as the first feature detection, an intermediate output of the model;
inputting the first feature detection into one or more of an additional model or a portion of the model;
receiving, from the one or more of the additional model or the portion of the model, a reconstructed input; and
determining a second difference between the reconstructed input and the portion of the sensor data,
wherein determining the annotated data comprises determining, using a statistical model, the annotated low variance region associated with the first feature detection.
18. The one or more non-transitory computer-readable media as claim 17 recites, wherein altering the one or more parameters is further based at least in part on the second difference.
19. The one or more non-transitory computer-readable media as claim 17 recites, wherein:
the model is a neural network,
the operations further comprising determining an additional high variance output, and
further wherein altering the one or more parameters comprises training the model end-to-end based at least in part on the low variance output, the high variance output, and the additional high variance output.
20. The one or more non-transitory computer-readable media as claim 16 recites, wherein the first feature detection comprises head detection and the second feature detection comprises a pedestrian detection.