Rare instance classifiers
In some implementations, an image classification system of an autonomous or semi-autonomous vehicle is capable of improving multi-object classification by reducing repeated incorrect classification of objects that are considered rarely occurring objects. The system can include a common instance classifier that is trained to identify and recognize general objects (e.g., commonly occurring objects and rarely occurring objects) as belonging to specified object categories, and a rare instance classifier that is trained to compute one or more rarity scores representing likelihoods that an input image is correctly classified by the common instance classifier. The output of the rare instance classifier can be used to adjust the classification output of the common instance classifier such that the likelihood of input images being incorrectly classified is reduced.
1. A method of training a rare instance neural network, the method comprising:
receiving a plurality of labeled training images;
identifying a subset of the labeled training images that are likely to be misclassified by a common instance neural network that has been trained on the labeled training images to classify images into object categories, wherein the identifying comprises at least one of:
processing the labeled training images using the trained common instance neural network to identify labeled training images that are misclassified by the common instance neural network, or
identifying labeled training images that depict objects of object types that occur rarely in the labeled training images;
generating, from the subset of the labeled training images, training data for a rare instance neural network, the rare instance neural network being configured to process an input image to generate a rarity output that includes a rarity score that represents a likelihood that the input image will be incorrectly classified by the trained common instance neural network; and
training the rare instance neural network on the training data to generate rarity scores for the subset of labeled training images that indicate that the subset of labeled training images are likely to be incorrectly classified by the common instance neural network.
2. The method of claim 1 , wherein the rare instance neural network has fewer parameters than the common instance neural network.
3. The method of claim 1 , wherein generating the training data comprises:
generating transformed images from one or more of the labeled training images.
4. The method of claim 1 , wherein training the rare instance neural network further comprises:
training the rare instance neural network to generate, for the subset of labeled training images, rarity scores that satisfy a predetermined threshold.
5. The method of claim 4 , wherein training further comprises:
training the rare instance neural network to generate, for labeled training images that are not in the subset of labeled training images, rarity scores that do not satisfy the predetermined threshold.
6. A system comprising:
one or more computers; and
one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:
receiving a plurality of labeled training images;
identifying a subset of the labeled training images that are likely to be misclassified by a common instance neural network that has been trained on the labeled training images to classify images into object categories, wherein the identifying comprises at least one of:
processing the labeled training images using the trained common instance neural network to identify labeled training images that are misclassified by the common instance neural network, or
identifying labeled training images that depict objects of object types that occur rarely in the labeled training images;
generating, from the subset of the labeled training images, training data for a rare instance neural network, the rare instance neural network being configured to process an input image to generate a rarity output that includes a rarity score that represents a likelihood that the input image will be incorrectly classified by the trained common instance neural network; and
training the rare instance neural network on the training data to generate rarity scores for the subset of labeled training images that indicate that the subset of labeled training images are likely to be incorrectly classified by the common instance neural network.
7. The system of claim 6 , wherein the rare instance neural network has fewer parameters than the common instance neural network.
8. The system of claim 6 , wherein generating the training data comprises:
generating transformed images from one or more of the labeled training images.
9. The system of claim 6 , wherein training the rare instance neural network further comprises:
training the rare instance neural network to generate, for the subset of labeled training images, rarity scores that satisfy a predetermined threshold.
10. The system of claim 9 , wherein training further comprises:
training the rare instance neural network to generate, for labeled training images that are not in the subset of labeled training images, rarity scores that do not satisfy the predetermined threshold.
11. A non-transitory computer-readable storage device encoded with computer program instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:
receiving a plurality of labeled training images;
identifying a subset of the labeled training images that are likely to be misclassified by a common instance neural network that has been trained on the labeled training images to classify images into object categories, wherein the identifying comprises at least one of:
processing the labeled training images using the trained common instance neural network to identify labeled training images that are misclassified by the common instance neural network, or
identifying labeled training images that depict objects of object types that occur rarely in the labeled training images;
generating, from the subset of the labeled training images, training data for a rare instance neural network, the rare instance neural network being configured to process an input image to generate a rarity output that includes a rarity score that represents a likelihood that the input image will be incorrectly classified by the trained common instance neural network; and
training the rare instance neural network on the training data to generate rarity scores for the subset of labeled training images that indicate that the subset of labeled training images are likely to be incorrectly classified by the common instance neural network.
12. The device of claim 11 , wherein the rare instance neural network has fewer parameters than the common instance neural network.
13. The device of claim 11 , wherein generating the training data comprises:
generating transformed images from one or more of the labeled training images.
14. The device of claim 11 , wherein training the rare instance neural network further comprises:
training the rare instance neural network to generate, for the subset of labeled training images, rarity scores that satisfy a predetermined threshold.