Divide BBOX_TRANSFORM into 2 ops.
authorXusong Wang <xusongw@google.com>
Thu, 4 Oct 2018 22:03:13 +0000 (15:03 -0700)
committerPrzemyslaw Szczepaniak <pszczepaniak@google.com>
Wed, 10 Oct 2018 17:00:48 +0000 (17:00 +0000)
Divide BBOX_TRANSFORM op into 2 ops
- AXIS_ALIGNED_BBOX_TRANSFORM
- ROTATED_BBOX_TRANSFORM

Rotated bounding boxes use different tensor format than axis-aligned
bounding boxes, and it would be less confusing if they were represented
by a separate operator code.

Bug: 113562630

Test: NeuralNetworksTest_static
Test: VtsHalNeuralnetworksV1_xTargetTest with sample driver
Change-Id: Ie08f2e0d0da77f6750766a394969653478d054d5
Merged-In: Ie08f2e0d0da77f6750766a394969653478d054d5
(cherry picked from commit d2bae1c268f8ef8e313441450e226d68ab515d9a)

neuralnetworks/1.2/types.hal

index 43771316d8a877e739435701fd69584d4458baa9..bed1d5cad90b5da76ab0e4e8d4d7f2ae5f790e60 100644 (file)
@@ -30,7 +30,7 @@ enum OperationType : @1.1::OperationType {
     ARGMAX = 38,
     ARGMIN = 39,
     PAD_V2 = 40,
-    BBOX_TRANSFORM = 41,
+    AXIS_ALIGNED_BBOX_TRANSFORM = 41,
     BIDIRECTIONAL_SEQUENCE_LSTM = 42,
     BIDIRECTIONAL_SEQUENCE_RNN = 43,
     BOX_WITH_NMS_LIMIT = 44,
@@ -76,6 +76,7 @@ enum OperationType : @1.1::OperationType {
     TRANSPOSE_CONV_2D = 84,
     UNIDIRECTIONAL_SEQUENCE_LSTM = 85,
     UNIDIRECTIONAL_SEQUENCE_RNN = 86,
+    ROTATED_BBOX_TRANSFORM = 87,
 };
 
 /**