Find Contours - Find Contours - 2026.1 English

The findcontours function detects object boundaries in a binary (8-bit) image using a chain-code contour tracer. Contour vertices are written to points_packed as 32-bit words; contour_offsets records where each contour begins and ends in that array, and num_contours returns the total number of contours found. The implementation is in L1/include/imgproc/xf_findcontours.hpp.

Output Point Packing

Each contour vertex is stored as one 32-bit word in points_packed:

• Bits [15:0] hold the x coordinate (16-bit unsigned).
• Bits [31:16] hold the y coordinate (16-bit unsigned).

All contours are stored back-to-back in points_packed. The contour_offsets array (length MAX_CONTOURS + 1) indexes into that flat array:

• contour_offsets[0] is always 0.
• For c >= 1, contour_offsets[c] is the exclusive end index for contour c - 1 (the number of 32-bit words consumed by contours 0 through c - 1).

Contour k therefore spans points_packed[contour_offsets[k]] through points_packed[contour_offsets[k + 1] - 1]. The number of points in contour k is contour_offsets[k + 1] - contour_offsets[k].

API Syntax

template <int SRC_T,
          int INPUT_PTR_WIDTH,
          int OUTPUT_PTR_WIDTH,
          int MAX_H,
          int MAX_W,
          int MAX_TOTAL_POINTS,
          int MAX_CONTOURS,
          int NPPCX = 1,
          int XFCVDEPTH_IN_MAT = 2>
void findcontours(xf::cv::Mat<SRC_T, MAX_H, MAX_W, NPPCX, XFCVDEPTH_IN_MAT>& _src,
                  int rows,
                  int cols,
                  ap_uint<OUTPUT_PTR_WIDTH>* points_packed,
                  ap_uint<OUTPUT_PTR_WIDTH>* contour_offsets,
                  ap_uint<OUTPUT_PTR_WIDTH>& num_contours);

Parameter Descriptions

The following table describes the template parameters.

The following table describes the function parameters.

Deviation from OpenCV

API and parameters

• There are no mode (for example RETR_EXTERNAL, RETR_TREE) or method (for example CHAIN_APPROX_SIMPLE, CHAIN_APPROX_NONE) arguments. Retrieval and chain encoding behavior are fixed in hardware.
• There is no hierarchy output and no optional offset argument.
• Outputs are written to points_packed, contour_offsets, and num_contours instead of an array of point lists.

Contour representation

• OpenCV returns each contour as a sequence of cv::Point values (full chain or approximated, depending on method).
• This kernel stores vertices only when the 8-connected chain direction changes, which compresses the boundary similarly in spirit to CHAIN_APPROX_SIMPLE but is not guaranteed to match OpenCV point for point.
• Each vertex is one 32-bit word y[31:16] | x[15:0] (see Output Point Packing above), not a cv::Point object.

Capacity and input

• MAX_CONTOURS and MAX_TOTAL_POINTS are compile-time limits. When either limit is reached, additional contours or points are not written.
• Input must be a single-channel 8-bit binary image (XF_8UC1).

Resource Utilization

The following table summarizes resource utilization for a 1200 x 800, single-channel (XF_8UC1) image, as generated in the Vitis HLS 2025.2 tool for Versal AI Edge at 300 MHz.

Performance Estimate

The following table summarizes performance for the same configuration (1200 x 800, XF_8UC1, 300 MHz).