Auto labeler
Aspects of the disclosure relate to training a labeling model to automatically generate labels for objects detected in a vehicle's environment. In this regard, one or more computing devices may receive sensor data corresponding to a series of frames perceived by the vehicle, each frame being captured at a different time point during a trip of the vehicle. The computing devices may also receive bounding boxes generated by a first labeling model for objects detected in the series of frames. The computing devices may receive user inputs including an adjustment to at least one of the bounding boxes, the adjustment corrects a displacement of the at least one of the bounding boxes caused by a sensing inaccuracy. The computing devices may train a second labeling model using the sensor data, the bounding boxes, and the adjustment to increase accuracy of the second labeling model when automatically generating bounding boxes.
1. A method of training a labeling model, comprising:
receiving, by one or more computing devices from one or more sensors of a vehicle, sensor data corresponding to a series of frames perceived by the vehicle, each frame being captured at a different time point during a trip of the vehicle;
receiving, by the one or more computing devices, bounding boxes generated by a first labeling model for objects detected in the series of frames, wherein the first labeling model is configured to automatically select a set of key frames among the series of frames, the set of key frames being frames in which a particular object detected in the set of key frames can be determined with errors within a predetermined threshold;
receiving, by the one or more computing devices, the set of key frames automatically selected by the first labeling model;
selecting, by the one or more computing devices, sensor data and bounding boxes corresponding to the set of key frames; and
training, by the one or more computing devices, a second labeling model using the selected sensor data and the selected bounding boxes to increase accuracy of the second labeling model when automatically generating bounding boxes.
2. The method of claim 1 , further comprising, receiving, by the one or more computing devices, one or more user inputs including at least one adjustment to at least one of the received bounding boxes, the adjustment corrects a displacement of the at least one of the received bounding boxes caused by a sensing inaccuracy, and wherein the training further uses the at least one adjustment.
3. The method of claim 2 , further comprising:
generating, by the one or more computing devices, at least one adjusted bounding box based on the at least one adjustment, wherein training the second labeling model is further based on the at least one adjusted bounding box.
4. The method of claim 2 , wherein the sensing inaccuracy results in a stationary object appearing to move between frames when the at least one bounding box of a first frame of the series of frames is compared to a respective bounding box of a second frame of the series of frames.
5. The method of claim 1 , wherein the sensing inaccuracy results in a moving object appearing to have a jittering trajectory when the at least one of the bounding boxes of a first frame of the series of frames is compared to a respective bounding box of a second frame of the series of frames.
6. The method of claim 1 , wherein the first labeling model is configured to automatically generate a position of the vehicle for each frame of the series of frames, and to automatically generate a trajectory of the vehicle based on the position for each frame of the series of frames.
7. The method of claim 6 , further comprising:
receiving, by the one or more computing devices, the trajectory of the vehicle automatically generated by the first labeling model;
receiving, by the one or more computing devices, one or more user inputs including at least one adjustment to the trajectory, the adjustment to the trajectory reduces jittering of the trajectory; and
training, by the one or more computing devices, the second labeling model using the sensor data, the trajectory, and the adjustment to the trajectory to automatically generate a smooth trajectory for the vehicle.
8. The method of claim 7 , further comprising generating, by the one or more computing devices, using the selected sensor data corresponding to the set of key frames, an interpolated trajectory for the particular object.
9. The method of claim 8 , further comprising presenting, by the one or more computing devices, the selected sensor data, selected bounding boxes, and the interpolated trajectory for review by a user.
10. The method of claim 1 , further comprising categorizing, by the one or more computing devices, one or more areas detected in the sensor data as no-label zones, wherein the no-label zones are excluded from the training.
11. The method of claim 1 , further comprising:
sending, by the one or more computing devices, the trained second labeling model to one or more computing devices of the vehicle for use onboard the vehicle.
12. A method of training a labeling model, comprising:
receiving, by one or more computing devices from one or more sensors of a vehicle, sensor data corresponding to a series of frames perceived by the vehicle, each frame being captured at a different time point during a trip of the vehicle;
receiving, by the one or more computing devices, bounding boxes generated by a first labeling model for objects detected in the series of frames, wherein the first labeling model is configured to automatically generate a position of the vehicle for each frame of the series of frames, and to automatically generate a trajectory of the vehicle based on the position for each frame of the series of frames;
receiving, by the one or more computing devices, the trajectory of the vehicle automatically generated by the first labeling model;
receiving, by the one or more computing devices, one or more user inputs including at least one adjustment to the trajectory, the adjustment to the trajectory reduces jittering of the trajectory; and
training, by the one or more computing devices, a second labeling model using the sensor data, the bounding boxes, the trajectory, and the adjustment to the trajectory to automatically generate a smooth trajectory for the vehicle.
13. The method of claim 12 , further comprising, receiving, by the one or more computing devices, one or more user inputs including at least one adjustment to at least one of the received bounding boxes, the adjustment corrects a displacement of the at least one of the received bounding boxes caused by a sensing inaccuracy, and wherein the training further uses the at least one adjustment.
14. The method of claim 13 , further comprising:
generating, by the one or more computing devices, at least one adjusted bounding box based on the at least one adjustment, wherein training the second labeling model is further based on the at least one adjusted bounding box.
15. The method of claim 14 , wherein the sensing inaccuracy results in a stationary object appearing to move between frames when the at least one bounding box of a first frame of the series of frames is compared to a respective bounding box of a second frame of the series of frames.
16. The method of claim 14 , wherein the sensing inaccuracy results in a moving object appearing to have a jittering trajectory when the at least one of the bounding boxes of a first frame of the series of frames is compared to a respective bounding box of a second frame of the series of frames.
17. The method of claim 16 , wherein the jittering trajectory results from bounding boxes for the moving object including different sets of points in consecutive frames which causes a displacement of the bounding boxes for the moving objects in the consecutive frames.
18. The method of claim 12 , wherein the jittering of the trajectory results as sensor inaccuracies causing displacements in the bounding boxes.
19. The method of claim 18 , wherein the sensor inaccuracies include sensor drift.
20. The method of claim 12 , wherein the user input changes one or more positions in the trajectory.