Hi, I am using the NVOF API in Cuda to get the flow vectors. I have followed this documentation.
I am using abgr format input and reference frames and I am receiving the vectors in S10.5 format which I am converting to float and doing some post processing to covert it into RGB displayable format. For some reason I am not able to get the correct visualization and getting some 4 divisions in the frame displayed. I have used some SDK code for post processing but it does not display as I expect it. I have run the same frames through AppOFCuda sample from the SDK and compared the visualization frame results. I believe I am getting incorrect vectors.
Does this function look correct way of calculating flow vectors?
void calculateFlow(uint8_t* frame1, uint8_t* frame2, uint8_t* vecframe, CUcontext cuContext, CUstream instream, CUstream outstream) {
// Create an instance of the API
API* nvofobj = new API(cuContext, instream, outstream);
// Initialize the input parameters
NV_OF_INIT_PARAMS initparams = { 0 };
initparams.width = W_BUFF;
initparams.height = H_BUFF;
initparams.inputBufferFormat = NV_OF_BUFFER_FORMAT_ABGR8;
initparams.mode = NV_OF_MODE_OPTICALFLOW;
initparams.outGridSize = (NV_OF_OUTPUT_VECTOR_GRID_SIZE)gridsize;
initparams.enableOutputCost = NV_OF_FALSE;
initparams.predDirection = NV_OF_PRED_DIRECTION_FORWARD;
initparams.perfLevel = NV_OF_PERF_LEVEL_SLOW;
initparams.enableExternalHints = NV_OF_FALSE;
initparams.enableRoi = NV_OF_FALSE;
initparams.hintGridSize = (NV_OF_HINT_VECTOR_GRID_SIZE)0;
NVOF_API_CALL(nvofobj->getAPI()->nvOFInit(nvofobj->getHandle(), &initparams));
// Buffer management
// Input frame buffer and uploading to GPU
NV_OF_BUFFER_DESCRIPTOR inbufferDesc;
inbufferDesc.width = W_BUFF;
inbufferDesc.height = H_BUFF;
inbufferDesc.bufferUsage = NV_OF_BUFFER_USAGE_INPUT;
inbufferDesc.bufferFormat = NV_OF_BUFFER_FORMAT_ABGR8;
NvOFCudaBuffer* inbuffer = new NvOFCudaBuffer(nvofobj, inbufferDesc, 1);
inbuffer->UploadData((void*)frame1);
// Reference frame buffer and uploading to GPU
NV_OF_BUFFER_DESCRIPTOR refbufferDesc;
refbufferDesc.width = W_BUFF;
refbufferDesc.height = H_BUFF;
refbufferDesc.bufferUsage = NV_OF_BUFFER_USAGE_INPUT;
refbufferDesc.bufferFormat = NV_OF_BUFFER_FORMAT_ABGR8;
NvOFCudaBuffer* refbuffer = new NvOFCudaBuffer(nvofobj, refbufferDesc, 1);
refbuffer->UploadData((void*)frame2);
// Pointer for storing the flow vectors
std::unique_ptr<NV_OF_FLOW_VECTOR[]> flowdata;
// calculate the dimensions of the output buffer
uint16_t outwidth = 0, outheight = 0;
outheight = H_BUFF / gridsize;
outwidth = W_BUFF / gridsize;
flowdata.reset(new NV_OF_FLOW_VECTOR[outwidth * outheight]);
// Output buffer
NV_OF_BUFFER_DESCRIPTOR outbufferDesc;
outbufferDesc.width = outwidth;
outbufferDesc.height = outheight;
outbufferDesc.bufferUsage = NV_OF_BUFFER_USAGE_OUTPUT;
outbufferDesc.bufferFormat = NV_OF_BUFFER_FORMAT_ABGR8;
NvOFCudaBuffer* outbuffer = new NvOFCudaBuffer(nvofobj, outbufferDesc, 1);
NV_OF_EXECUTE_INPUT_PARAMS inparams;
memset(&inparams, 0, sizeof(NV_OF_EXECUTE_INPUT_PARAMS));
inparams.inputFrame = inbuffer->getOFBufferHandle();
inparams.referenceFrame = refbuffer->getOFBufferHandle();
inparams.externalHints = nullptr;
inparams.disableTemporalHints = NV_OF_FALSE;
inparams.hPrivData = nullptr;
inparams.numRois = 0;
inparams.padding = 0;
inparams.padding2 = 0;
NV_OF_EXECUTE_OUTPUT_PARAMS outparams;
memset(&outparams, 0, sizeof(NV_OF_EXECUTE_OUTPUT_PARAMS));
outparams.bwdOutputBuffer = nullptr;
outparams.bwdOutputCostBuffer = nullptr;
outparams.globalFlowBuffer = nullptr;
outparams.hPrivData = nullptr;
outparams.outputBuffer = outbuffer->getOFBufferHandle();
outparams.outputCostBuffer = nullptr;
// Run Optical Flow
nvofobj->getAPI()->nvOFExecute(nvofobj->getHandle(), &inparams, &outparams);
// Will contain the flow vectors in direction x and y for each pixel
outbuffer->DownloadData((void*)flowdata.get());
postProcessVectors((const NV_OF_FLOW_VECTOR*)flowdata.get(), (uint8_t*)vecframe, outwidth, outheight);
flowdata.reset();
// Destroy the buffers
delete inbuffer;
delete refbuffer;
delete outbuffer;
// Destroy NVOF session
nvofobj->getAPI()->nvOFDestroy(nvofobj->getHandle());
}
My resulting frame
