Machine learning techniques
Improved techniques for training a machine learning (ML) model are discussed herein. Training the ML model can be based on a subset of examples. In particular, the training can include identifying a reference region associated with an area of the image representing an object, and selecting, based at least in part on a first confidence score associated with a first bounding box, a first hard example for inclusion in the subset of examples. In some cases, the first confidence score and the first bounding box can be associated with a first portion of the feature map. Next, the training can include determining that a first degree of alignment of the first bounding box to the reference region is above a threshold degree of alignment, and in response, replacing the first hard example with a second hard example.
1. A computer-implemented method comprising:
receiving an image representing an object;
identifying a reference region indicative of an area representing the object in the image;
selecting a subset of portions of the image; and
training, based at least in part on the subset of portions, a machine-learning (ML) model to output a feature map, a portion of the feature map associated with classification information, region of interest (ROI) information, and confidence information,
wherein selecting the subset of portions of the image comprises:
determining a first portion of the feature map based at least in part on the first portion having first confidence information that meets or exceeds a threshold confidence; and
determining to include, based at least in part on ROI information associated with the first portion, a second portion of the feature map in the subset of portions of the image, wherein the second portion is different than the first portion.
2. The computer-implemented method as claim 1 recites, wherein:
determining to include the second portion is based at least in part on performing a non-maximal suppression (NMS) of the ROI information of the first portion with respect to the reference region; and
the second portion of the feature map is associated with an ROI suppressed under non-maximal suppression.
3. The computer-implemented method as claim 2 recites, further comprising:
associating, with the first portion, a highest confidence score.
4. The computer-implemented method as claim 1 recites, further comprising:
training the ML model to substantially simultaneously output multiple output ROIs corresponding to classifications based at least in part on the image, the classifications comprising one or more of a vehicle, a pedestrian, or a bicyclist.
5. The computer-implemented method as claim 1 recites, wherein:
the reference region comprises a first region and a second region, and
further wherein the second portion of the feature map comprises features outside of the first region.
6. The computer-implemented method as claim 5 recites, wherein the first region is associated with:
a central portion of the reference region; and
a positive identification of classification information.
7. The computer-implemented method as claim 1 recites, wherein determining to include the second portion is based at least in part on determining an amount of alignment as an intersection of the ROI information associated with the first portion with the area indicated by the reference region divided by a union of the ROI information associated with the first portion and the reference region.
8. The computer-implemented method as claim 7 recites, wherein determining to include the second portion is based at least in part on determining the amount of alignment meets or exceeds a threshold degree of alignment.
9. The computer-implemented method as claim 1 recites, wherein training the ML model comprises at least two stages, wherein:
a first stage of the at least two stages comprises:
providing, as first input to the ML model, a first batch of images;
receiving, from the ML model and based at least in part on the first batch of images,
an intermediate ROI associated with a first image of the first batch of images; and
identifying, as the reference region, the intermediate ROI; and
a second stage of the at least two stages comprises:
providing, as second input to the ML model, the first batch of images;
receiving from the ML model and based at least in part on the first batch of images, a second plurality of ROIs and a second plurality of confidence scores associated therewith;
selecting, for inclusion in a second subset of examples for training the ML model and based at least in part on a non-maximal suppression (NMS) technique,
a hard example for training the ML model from among portions of the image associated with the second plurality of ROIs; and
training the ML model based at least in part on the second subset of examples.
10. The computer-implemented method as claim 9 recites, wherein the second stage further comprises training the ML model using focal loss.
11. The computer-implemented method as claim 1 recites, wherein the ROI information comprises a bounding box.
12. The computer-implemented method as claim 1 recites, further comprising:
receiving a batch of images comprising a first predefined number of images that are associated with a first object classification and a second predefined number of images that are associated with a second object classification; and
training the ML model based at least in part on providing the batch of images as input to the ML model,
wherein the first predefined number of images and the second predefined number of images are based at least in part on confidence information associated with one or more features of the feature map.
13. A system comprising:
one or more processors; and
one or more non-transitory computer-readable media storing computer-executable instructions that, when executed, cause the one or more processors to perform operations comprising:
receiving an image representing an object;
identifying a reference region indicative of an area representing the object in the image;
selecting a subset of portions of the image; and
training, based at least in part on the subset of portions, a machine-learning (ML) model to output a feature map, a portion of the feature map associated with classification information, region of interest (ROI) information, and confidence information,
wherein selecting the subset of portions of the image comprises:
identifying a first portion of the feature map based at least in part on the first portion having first confidence information that meets or exceeds a threshold confidence;
comparing, as a comparison, ROI information associated with the first portion with the reference region; and
determining, as the subset of portions of the image and based at least in part on the comparison, a second portion of the feature map.
14. The system of claim 13 , wherein:
comparing the ROI comprises performing a non-maximal suppression (NMS) of the ROI information of the first portion with respect to the reference region; and
the second portion of the feature map is associated with an ROI suppressed under non-maximal suppression.
15. The system of claim 13 , wherein:
the reference region comprises a first region and a second region, and
further wherein the second portion of the feature map comprises features outside of the first region.
16. The system of claim 13 , wherein comparing comprises determining an amount of alignment as an intersection of the ROI information with the area indicated by the reference region divided by a union of the ROI information and the reference region.
17. The system of claim 13 , wherein training the ML model comprises at least two stages, wherein:
a first stage of the at least two stages comprises:
providing, as first input to the ML model, a first batch of images;
receiving, from the ML model and based at least in part on the first batch of images,
an intermediate ROI associated with a first image of the first batch of images; and
identifying, as the reference region, the intermediate ROI; and
a second stage of the at least two stages comprises:
providing, as second input to the ML model, the first batch of images;
receiving from the ML model and based at least in part on the first batch of images, a second plurality of ROIs and a second plurality of confidence scores associated therewith;
selecting, for inclusion in a second subset of examples for training the ML model and based at least in part on a non-maximal suppression (NMS) technique,
a hard example for training the ML model from among portions of the image associated with the second plurality of ROIs; and
training the ML model based at least in part on the second subset of examples.
18. One or more non-transitory computer-readable media storing instructions executable by a processor, wherein the instructions, when executed, cause the processor to perform operations comprising:
receiving an image representing an object;
identifying a reference region indicative of an area representing the object in the image;
selecting a subset of portions of the image; and
training, based at least in part on the subset of portions, a machine-learning (ML) model to output a feature map, a portion of the feature map associated with classification information, region of interest (ROI) information, and confidence information,
wherein selecting the subset of portions of the image comprises:
identifying a first portion of the feature map based at least in part on the first portion having first confidence information that meets or exceeds a threshold confidence;
comparing, as a comparison, ROI information associated with the first portion with the reference region; and
determining, as the subset of portions of the image and based at least in part on the comparison, a second portion of the feature map.
19. The one or more non-transitory computer-readable media of claim 18 , wherein:
comparing the ROI comprises performing a non-maximal suppression (NMS) of the ROI information of the first portion with respect to the reference region; and
the second portion of the feature map is associated with an ROI suppressed under non-maximal suppression.
20. The one or more non-transitory computer-readable media of claim 18 , wherein:
the reference region comprises a first region and a second region, and
further wherein the second portion of the feature map comprises features outside of the first region.