3D object detection
A method of training a 3D structure detector to detect 3D structure in 3D structure representation, the method comprising the following steps: receiving, at a trainable 3D structure detector, a set of training inputs, each training input comprising at least one 3D structure representation; the 3D structure detector determining, for each training input, a set of predicted 3D objects for the at least one 3D structure representation of that training input; and training the 3D structure detector to optimize a cost function, wherein the cost function penalizes deviation from an expected geometric relationship between the set of predicted 3D objects determined for each training in put.
1. A method of training a 3D structure detector to detect 3D structure in 3D structure representation, the method comprising the following steps:
receiving, at a trainable 3D structure detector, a set of training inputs, each training input comprising at least one 3D structure representation;
the 3D structure detector determining, for each training input, a set of predicted 3D objects for the at least one 3D structure representation of that training input; and
training the 3D structure detector to optimize a cost function, wherein the cost function penalizes deviation from an expected geometric relationship between the set of predicted 3D objects for the at least one 3D structure representation, determined for each training input.
2. The method of claim 1 , wherein each training input comprises at least two 3D structure representations, wherein the set of predicted 3D objects are predicted 3D objects for a common 3D structural element captured in the 3D structure representations of that training input.
3. A method according to claim 2 , wherein one or more of the 3D structure representations are selected from at least one temporal sequence of structure representations, the 3D objects corresponding to the common 3D structural element at different times.
4. A method according to claim 3 , wherein object tracking is applied, by an object tracker, to the at least one temporal sequence of structure representations, wherein the expected geometric relationship is determined for each of one or more of the training inputs based on tracking of the common 3D structural element by the object tracker.
5. A method according to claim 3 , wherein the at least one temporal sequence of structure representations is a sequence of images comprised in at least one video image.
6. A method according to claim 2 , wherein at least one of the training inputs comprises at least two 3D structure representations captured by different sensor units, wherein the expected geometric relationship between the predicted 3D objects is determined based on a known geometric relationship between the different sensor units, the predicted 3D objects corresponding to the common 3D structural element captured from the different sensor units.
7. A method according to claim 6 , wherein the known geometric relationship is determined by processing data captured by the different sensor units.
8. A method according to claim 1 , wherein the set of predicted 3D objects are predicted 3D objects for component parts of a 3D object captured in the at least one 3D structure representation, the cost function penalizing deviation from an expected geometric relationship between the component parts.
9. A method according to claim 1 , comprising a step of determining, for each of one or more of the training inputs, a size difference between the set of predicted 3D objects, wherein the cost function penalizes any deviation in the determined size difference from an expected size difference for each of the one or more training inputs.
10. A method according to claim 9 , wherein each training input comprises at least two 3D structure representations, wherein the set of predicted 3D objects are predicted 3D objects for a common 3D structural element captured in the 3D structure representations of that training input, and wherein the expected size difference is zero.
11. A method according to claim 1 , comprising a step of determining, for each of one or more of the training inputs, a position difference between the set of predicted 3D objects, wherein the cost function penalizes any deviation in the determined position difference from an expected position difference.
12. A method according to claim 4 , comprising a step of determining, for each of one or more of the training inputs, a position difference between the set of predicted 3D objects, wherein the cost function penalizes any deviation in the determined position difference from an expected position difference, wherein the expected position difference is determined based on the tracking of the common 3D structural element by the object tracker.
13. A method according to claim 1 , comprising a step of determining, for each of one or more of the training inputs, an orientation difference between the set of predicted 3D objects, wherein the cost function penalizes any deviation in the determined orientation difference from an expected orientation difference.
14. A method according to claim 4 , comprising a step of determining, for each of at least some of the training inputs, an orientation difference between the set of predicted 3D objects, wherein the cost function penalizes any deviation in the determined orientation difference from an expected orientation difference, wherein the expected orientation difference is determined based on the tracking of the common 3D structural element by the object tracker.
15. A method according to claim 8 , wherein the captured 3D object comprises a vehicle and the component parts comprise vehicle components; or
wherein the captured 3D object comprises a living being, and the component parts comprise body parts; or
wherein the captured 3D object comprises both a living being and a vehicle, and the component parts comprise one or more body parts and one or more vehicle components.
16. A method according to claim 1 , wherein a subset of the 3D structure representations have associated ground truth 3D object annotation data, and the cost function also penalizes deviations between the ground truth 3D object annotation data and the predicted 3D objects determined for those 3D structure representations by the 3D object detector.
17. A method according to claim 1 , wherein each 3D object is represented as a 3D boundary object.
18. A method according to claim 17 , wherein the at least one 3D structure representation received at the 3D structure detector have associated 2D structure annotation data, and the cost function also penalizes deviation between a predicted 2D boundary object determined for each of those 3D structure representations and the associated 2D structure annotation data, wherein the predicted 2D boundary object is determined by projecting the predicted 3D boundary object determined for that 3D structure representations into a 2D image plane.
19. Non-transitory media embodying a 3D object detector as a set of executable program instructions, the 3D object detector having been trained by:
receiving, at the 3D object detector, a set of training inputs, each training input comprising at least one 3D structure representation:
the 3D object detector determining, for each training input, a set of predicted 3D objects for the at least one 3D structure representation of that training input; and
training the 3D object detector to optimize a cost function, wherein the cost function penalizes deviation from an expected geometric relationship between the set of predicted 3D objects for the at least one 3D structure representation determined for each training input.
20. A processing system comprising: execution hardware configured to train a 3D structure detector to detect 3D structure in 3D structure representation, by:
receiving, at the 3D structure detector, a set of training inputs, each training input comprising at least one 3D structure representation:
the 3D structure detector determining, for each training input, a set of predicted 3D objects for the at least one 3D structure representation of that training input: and
training the 3D structure detector to optimize a cost function, wherein the cost function penalizes deviation from an expected geometric relationship between the set of predicted 3D objects for the at least one 3D structure representation determined for each training input.