Methods and apparatus localizing object(s) in vision data
An object localization network (OLN) can be used to localize object(s) (e.g., known and/or unknown object(s)) in an instance of vision data. Various implementations include detecting the localized object(s) based on the localization. Many implementations include processing the instance of vision data using the OLN to generate a objectness score (e.g., a centerness score) as well as an intersection of union (IoU) score for one or more proposed object locations in the instance of vision data. Object(s) can be localized in the instance of vision data based on the objectness scores and the IoU scores.
1 . A method implemented by one or more processors, the method comprising:
training an Object Localization Network (OLN) to detect one or more objects, in an instance of vision data, based on predicted object localization, wherein training the OLN comprises:
selecting a training instance including a training instance of vision data and one or more training annotations, where each training annotation in the one or more training annotations identifies a ground truth location of a corresponding object location of an object, in the training instance of vision data;
processing the training instance of vision data using a feature extraction network portion of the OLN to generate a set of training instance features;
processing the set of training instance features using a region proposal portion of the OLN to generate a set of proposed object locations in the training instance of vision data;
processing the set of proposed object locations using a region based portion of the OLN to generate one or more predicted annotations, where each predicted annotation in the one or more predicted annotations identifies a corresponding predicted location of a corresponding object location in the training instance of vision data;
generating a loss based on comparing the one or more predicted annotations and the one or more training annotations of the training instance of vision data; and
updating one or more portions of the OLN based on the generated loss;
wherein processing the set of training instance features using the region proposal portion of the OLN to generate the set of proposed object locations in the training instance of vision data comprises:
initializing a set of anchor boxes across the training instance of vision data;
processing the set of anchor boxes and the set of training instance features to generate the set of proposed object locations, wherein processing the set of anchor boxes and the set of training instances features to generate the set of proposed object locations comprises:
processing the set of anchor boxes and the set of training instance features using an objectness portion of the region proposal portion to generate a centerness score for each anchor box, in the set of anchor boxes, wherein the centerness score indicates a deviation from a corresponding ground truth object center; and
generating the set of proposed object locations based on the corresponding centerness scores.
2 . The method of claim 1 , wherein training the OLN further comprises:
selecting an additional training instance including an additional training instance of vision data and one or more additional training annotations, where each additional training annotation in the one or more additional training annotations identifies an additional ground truth location of an additional corresponding object location of an additional object, in the additional training instance of vision data;
processing the additional training instance of vision data using the feature extraction network portion of the OLN to generate an additional set of training instance features;
processing the additional set of training instance features using the region proposal portion of the OLN to generate an additional set of proposed object locations in the additional training instance of vision data;
processing the additional set of proposed object locations using the region based portion of the OLN to generate one or more additional predicted annotations, where each additional predicted annotation in the one or more additional predicted annotations identifies a corresponding predicted location of a corresponding additional object location in the additional training instance of vision data;
generating an additional loss based on comparing the one or more additional predicted annotations and the one or more additional training annotations of the additional training instance of vision data; and
updating one or more additional portions of the OLN based on the generated additional loss.
3 . The method of claim 1 , wherein the feature extraction network portion of the OLN includes one or more feature pyramid connections, and wherein the set of training instance features include one or more feature maps.
4 . The method of claim 1 , further comprising:
receiving a given instance of vision data;
detecting one or more given objects in the given instance of vision data based on processing the given instance of vision data using the trained OLN.
5 . The method of claim 4 , wherein detecting the one or more given objects in the given instance of vision data based on processing the given instance of vision data using the trained OLN comprises:
processing the given instance of vision data using the feature extraction network portion of the OLN to generate a set of given instance features;
processing the set of given instance features using the region proposal portion of the OLN to generate a set of given proposed object locations in the given instance of vision data;
processing the set of given proposed object locations using the region based portion of the OLN to generate one or more given predicted annotations, where each given predicted annotation in the one or more given predicted annotations identifies a corresponding predicted location of a corresponding object location in the given instance of vision data; and
generating the one or more given objects in the given instance of vision data based on the set of given proposed object locations and the one or more given predicted annotations.
6 . The method of claim 1 , wherein processing the set of proposed object locations using the region based portion of the OLN to generate the one or more predicted annotations, where each predicted annotation in the one or more predicted annotations identifies the corresponding predicted location of a corresponding object location in the training instance of vision data comprises:
processing the set of proposed object locations using a box refinement portion of the region based portion of the OLN to generate a set of refined candidate object locations in the training instance of vision data;
processing the set of proposed object locations using an intersection of union (IoU) portion of the region based portion of the OLN to generate an IoU score for each proposed object location in the set of proposed object locations, where the IoU score indicates an intersection of union between the corresponding proposed object location and a corresponding training annotation identifying the ground truth location of the corresponding object location of the object; and
generating the one or more predicted annotations based on the set of refined candidate object locations and the corresponding IoU scores.
7 . The method of claim 1 , wherein the region based portion of the OLN is a region based convolutional neural network (R-CNN).
8 . The method of claim 7 , wherein the R-CNN is a faster R-CNN.
9 . The method of claim 1 , wherein each training annotation of the one or more training annotations includes a training bounding box identifying the corresponding object location of an object in the training instance of vision data.
10 . The method of claim 1 , wherein each predicted annotation of the one or more predicted annotations includes a predicted bounding box identifying the corresponding object location of a predicted object in the training instance of vision data.
11 . A method implemented by one or more processors, the method comprising:
identifying an instance of vision data including one or more objects;
detecting the one or more objects in the instance of vision data, wherein detecting the one or more objects comprises:
processing the instance of vision data using an Object Localization Network (OLN) to generate one or more predicted annotations, where each predicted annotation in the one or more predicted annotations identifies a corresponding predicted location of a corresponding object, of the one or more objects in the instance of vision data, wherein processing the instance of vision data using the OLN to generate the one or more predicted annotations comprises:
processing the instance of vision data using a box refinement portion of the OLN to generate a set of proposed object locations in the instance of vision data;
for each proposed object location in the set of proposed object locations, processing the proposed object location using a centerness portion of the OLN to generate a centerness score corresponding to the proposed object location and indicating a predicted deviation of the proposed object location from a ground truth representation of the location of the object;
selecting a subset of proposed object locations from the set of proposed object locations based on the corresponding centerness scores;
processing the subset of the proposed object locations using an additional box refinement portion of the OLN to generate a set of refined object locations;
for each refined object location in the set of refined object locations, processing the refined object location using an intersection of union (IoU) portion of the OLN to generate a corresponding IoU score indicating a predicted intersection of union between the refined object location and the ground truth representation of the object;
ranking the set of refined object locations, where the rank of each refined object location, in the set of refined object locations, is based on the corresponding centerness score and the corresponding IoU score; and
detecting the one or more objects based on the ranking of the set of refined object locations.
12 . The method of claim 11 , wherein at least one of the one or more objects detected in the instance of vision data are captured in one or more training instances used to train the OLN.
13 . The method of claim 11 , wherein at least one of the one or more objects detected in the instance of vision data were not captured in one or more training instances used to train the OLN.