Non-Maximum Suppression

group NVCV_C_ALGORITHM_NON_MAXIMUM_SUPPRESSION

Unnamed Group

NVCVStatus cvcudaNonMaximumSuppressionSubmit(NVCVOperatorHandle handle, cudaStream_t stream, NVCVTensorHandle in, NVCVTensorHandle out, NVCVTensorHandle scores, float scoreThreshold, float iouThreshold)

Executes the Non-Maximum Suppression operation on the given cuda stream.

Non-Maximum-Suppression operation finds the set of non-overlapping bounding boxes (bboxes), for a given set of bboxes and scores, based on a score threshold and Intersection-over-Union (IoU) threshold. First, all input bboxes with scores less than the score threshold are discarded. Then, all input bboxes with sufficient overlap and lower score than another input bbox are discarded. The overlap is calculated via IoU fraction, i.e. the intersection area divided by the area of the union, where sufficient overlap means greater than IoU threshold.

Limitations:

Input: Data Layout: [NW] Channel count: [4]

Data Type	Allowed
8bit Unsigned	No
8bit Signed	No
16bit Unsigned	No
16bit Signed	Yes
32bit Unsigned	No
32bit Signed	No
32bit Float	No
64bit Float	No

Output: Data Layout: [NW] Channel count: [1]

Data Type	Allowed
8bit Unsigned	Yes
8bit Signed	No
16bit Unsigned	No
16bit Signed	No
32bit Unsigned	No
32bit Signed	No
32bit Float	No
64bit Float	No

Input/Output dependency

 Property      |  Input == Output
-------------- | -------------
 Data Layout   | Yes
 Data Type     | No
 Batches (N)   | Yes
 Bboxes (W)    | Yes
 Channels      | No

Parameters

• handle – [in] Handle to the operator.

  • Must not be NULL.

• stream – [in] Handle to a valid CUDA stream.

• in – [in] Input tensor, in[i, j] is the set of input bbox proposals for an image where i ranges from 0 to batch-1, j ranges from 0 to number of bbox proposals, and the data type being short4 for (x=x, y=y, z=width, w=height) anchored at the top-left of the bounding box area

  • Must have data type 4S16 or S16, in case of S16 last shape must be 4

  • Must have rank 2 or 3, in case of 2 last shape must be 1 and data type must be 4S16

• out – [out] Output tensor, out[i, j] is the output boolean mask marking the selected bboxes for an image where i ranges from 0 to batch-1, j ranges from 0 to the number of bbox proposals, and the data type being uint8_t marking selected bboxes as ones and discarded bboxes as zeros

  • Must have data type U8

  • Must have rank 2 or 3, in case of 3 last shape must be 1

• scores – [in] Input tensor, scores[i, j] are the associated scores for each bounding box proposal in in considered during the reduce operation

  • Must have data type F32

  • Must have rank 2 or 3, in case of 3 last shape must be 1

• scoreThreshold – [in] Minimum score an input bbox proposal need to have to be kept

• iouThreshold – [in] Maximum overlap between bbox proposals covering the same effective image region as calculated by Intersection-over-Union (IoU) fraction

Return values

• NVCV_ERROR_INVALID_ARGUMENT – Some parameter is outside valid range.

• NVCV_ERROR_INTERNAL – Internal error in the operator, invalid types passed in.

• NVCV_SUCCESS – Operation executed successfully.

Functions

NVCVStatus cvcudaNonMaximumSuppressionCreate(NVCVOperatorHandle *handle)

Constructs and an instance of the Non-Maximum-Suppression operator.

Parameters

handle – [out] Where the image instance handle will be written to.

• Must not be NULL.

Return values

• NVCV_ERROR_INVALID_ARGUMENT – Handle is null.

• NVCV_ERROR_OUT_OF_MEMORY – Not enough memory to create the operator.

• NVCV_SUCCESS – Operation executed successfully.