Runs object detection for the input stream specified with -i option. By default INT8 quantized ssd_mobilenet_v1_coco model is used. Use -detect_type yolo with -m_detect to change model type and point to Yolov4 model location.
-dconf option specifies confidence threshold: only results with the higher confidence then the threshold will be posted. -b options sets batch size.
In order to evaluate performance on multiple input streams, use -c option to emulate multiple inputs. Variate -infer to adjust number of inference instances (which will handle incoming requests from all of the input streams) and -nireq to adjust number of requests each inference instance can handle in parallel.
- --help
-
Print help
- -codec 264|265
-
Use this codec to decode the video (default: 264)
- -c channels
-
Number of input channels (default: 1). This option allows to emulate processing of multiple input streams (channels) in the same application. With this option input stream will be used specified number of times in parallel decoding+inference sessions. Each channel will be handled by dedicated decoder. Mind however that inference instances work independently and might process requests from different channels. Number of inference instances and how many requests can be processed by each are controlled by -infer and -nireq options respectively.
- -m_detect model
-
xml model file name with absolute path, no .xml needed (default: /opt/intel/samples/models/ssd_mobilenet_v1_coco_INT8/ssd_mobilenet_v1_coco)
- -detect_type ssd|yolo
-
Detection model type (default: ssd)
- -dshape_w width
-
Detection model input reshape width
-
ssd default: 300
-
yolo default: 416
-
- -dshape_h height
-
Detection model input reshape height
-
ssd default: 300
-
yolo default: 416
-
- -dconf threshold
-
Minimum detection output confidence, range [0-1] (default: 0.8)
- -b batch_number
-
Batch number in the inference model (default: 1)
- -infer inference_instances_number
-
Number of independent inference instances to process incoming inference requests in parallel (default: 1). Each inference instance can accept few requests for parallel processing, number of such requests is controlled by -nireq option.
- -nireq req_number
-
Number of requests (default: 1)
- -r vp_ratio
-
Ratio of decoded frames to inference frames (default: 1), =2 means doing inference every other frame. "vp" stands for video processing meaning scaling operation to get input for the inference.
- -scale hq|fast_inplace|fast
-
Scaling mode to use:
-
hq - run scaling on EUs thru rcs or ccs depending on the platform
-
fast_inplace - run scaling with affinity to decoding, i.e. on the same vcs; scaling is done thru SFC
-
fast - run scaling via vecs, scaling is done thru SFC (this is default)
-
- -t seconds
-
How many seconds this app should run. If this option is specified input stream is being used in a loop.
- -d
-
Dump inference input (= vp output).
- -p
-
Toggle performance mode, don’t dump the csv results.
- -va_share
-
Share surfaces between media and inferece. This is performance optimization option and is strongly recommended.
- output.csv
-
This is major output containing detected objects label IDs and bounding box coordinates. Output can be disabled with -p option. File format:
channel#, frame#, object#, left, top, right, bottom, id, probability
- VPOut_%d.300x300.rgbp
-
Rough output in RGBP of the inference input (these are frames scaled per model requirements). Output is produced per each channel (mind %d in the file name pattern) only if -d option is specified on a command line.