@@ -113,7 +113,7 @@ Output (format, keypoint index ordering, etc.) in [doc/output.md](doc/output.md)
## Speeding Up OpenPose and Benchmark
Check the OpenPose Benchmark and some hints to speed up OpenPose on [doc/faq.md#speed-up-and-benchmark](doc/faq.md#speed-up-and-benchmark).
Check the OpenPose Benchmark as well as some hints to speed up and/or reduce the memory requirements for OpenPose on [doc/faq.md#speed-up-memory-reduction-and-benchmark](doc/faq.md#speed-up-memory-reduction-and-benchmark).
Note: This example will assume that the target are 3 Flir/Point Grey cameras, but it can be generalized to any camera model.
1. Distortion and intrinsic parameter calibration:
1. Run OpenPose and save images for your desired camera. Use a grid (chessboard) pattern and move around all the image area. I.e., cover several distances, and within each distance, cover all parts of the image view (all corners and center). You should get at least 100-150 images for a good calibration. Note: we recommend saving the images in PNG format (default behavior) in order to improve calibration quality.
2. Flir camera calibration: Add the flags `--flir_camera --flir_camera_index 0` (or the desired flir camera index) to the webcam command.
3. Calibration from video sequence: Add the flag `--video {video_path}` to the webcam command.
4. Any other camera brand: Simply save your images in {intrinsic_images_folder_path}, file names are not relevant.
1. Run OpenPose and save images for your desired camera. Use a grid (chessboard) pattern and move around all the image area.
1. Quality tips:
1. Keep the same orientation of the chessboard, i.e., do not rotate it circularly more than ~15-30 degress. Our algorithm assumes that the origin is the corner at the top left, so rotating the chessboard circularly will change this origin across frames, resulting in many frames being rejected for the final calibration, i.e., lower calibration accuracy.
2. You should get at least 100-150 images for a good calibration.
3. Cover several distances, and within each distance, cover all parts of the image view (all corners and center).
4. Save the images in PNG format (default behavior) in order to improve calibration quality.
5. Calibration of about 100 images takes about 3 minutes.
@@ -134,6 +134,7 @@ Each flag is divided into flag name, default value, and description.
- DEFINE_int32(logging_level, 3, "The logging level. Integer in the range [0, 255]. 0 will output any log() message, while 255 will not output any. Current OpenPose library messages are in the range 0-4: 1 for low priority messages and 4 for important ones.");
- DEFINE_bool(disable_multi_thread, false, "It would slightly reduce the frame rate in order to highly reduce the lag. Mainly useful for 1) Cases where it is needed a low latency (e.g. webcam in real-time scenarios with low-range GPU devices); and 2) Debugging OpenPose when it is crashing to locate the error.");
- DEFINE_int32(profile_speed, 1000, "If PROFILER_ENABLED was set in CMake or Makefile.config files, OpenPose will show some runtime statistics at this frame number.");
2. Producer
- DEFINE_int32(camera, -1, "The camera index for cv::VideoCapture. Integer in the range [0, 9]. Select a negative number (by default), to auto-detect and open the first available camera.");
- DEFINE_string(camera_resolution, "-1x-1", "Set the camera resolution (either `--camera` or `--flir_camera`). `-1x-1` will use the default 1280x720 for `--camera`, or the maximum flir camera resolution available for `--flir_camera`");
...
...
@@ -151,12 +152,13 @@ Each flag is divided into flag name, default value, and description.
- DEFINE_bool(process_real_time, false, "Enable to keep the original source frame rate (e.g. for video). If the processing time is too long, it will skip frames. If it is too fast, it will slow it down.");
- DEFINE_string(camera_parameter_folder, "models/cameraParameters/flir/", "String with the folder where the camera parameters are located.");
- DEFINE_bool(frame_keep_distortion, false, "If false (default), it will undistortionate the image based on the `camera_parameter_folder` camera parameters; if true, it will not undistortionate, i.e., it will leave it as it is.");
3. OpenPose
- DEFINE_string(model_folder, "models/", "Folder path (absolute or relative) where the models (pose, face, ...) are located.");
- DEFINE_string(output_resolution, "-1x-1", "The image resolution (display and output). Use \"-1x-1\" to force the program to use the input image resolution.");
- DEFINE_int32(num_gpu, -1, "The number of GPU devices to use. If negative, it will use all the available GPUs in your machine.");
- DEFINE_int32(keypoint_scale, 0, "Scaling of the (x,y) coordinates of the final pose data array, i.e. the scale of the (x,y) coordinates that will be saved with the `write_keypoint` & `write_keypoint_json` flags. Select `0` to scale it to the original source resolution, `1`to scale it to the net output size (set with `net_resolution`), `2` to scale it to the final output size (set with `resolution`), `3` to scale it in the range [0,1], and 4 for range [-1,1]. Non related with `scale_number` and `scale_gap`.");
- DEFINE_int32(keypoint_scale, 0, "Scaling of the (x,y) coordinates of the final pose data array, i.e. the scale of the (x,y) coordinates that will be saved with the `write_keypoint` & `write_keypoint_json` flags. Select `0` to scale it to the original source resolution; `1`to scale it to the net output size (set with `net_resolution`); `2` to scale it to the final output size (set with `resolution`); `3` to scale it in the range [0,1], where (0,0) would be the top-left corner of the image, and (1,1) the bottom-right one; and 4 for range [-1,1], where (-1,-1) would be the top-left corner of the image, and (1,1) the bottom-right one. Non related with `scale_number` and `scale_gap`.");
- DEFINE_int32(number_people_max, -1, "This parameter will limit the maximum number of people detected, by keeping the people with top scores. The score is based in person area over the image, body part score, as well as joint score (between each pair of connected body parts). Useful if you know the exact number of people in the scene, so it can remove false positives (if all the people have been detected. However, it might also include false negatives by removing very small or highly occluded people. -1 will keep them all.");
**Q: Low speed** - OpenPose is quite slow, is it normal? How can I speed it up?
**A**: Check the [OpenPose Benchmark](https://docs.google.com/spreadsheets/d/1-DynFGvoScvfWDA1P4jDInCkbD4lg0IKOYbXgEq0sK0/edit#gid=0) to discover the approximate speed of your graphics card. Some speed tips:
There are 2 alternatives to save the OpenPose output.
1. The `write_json` flag saves the people pose data using a custom JSON writer. Each JSON file has a `people` array of objects, where each object has:
1. An array `pose_keypoints_2d` containing the body part locations and detection confidence formatted as `x1,y1,c1,x2,y2,c2,...`. The coordinates `x` and `y` can be normalized to the range [0,1], [-1,1], [0, source size], [0, output size], etc., depending on the flag `keypoint_scale`, while `c` is the confidence score in the range [0,1].
1. An array `pose_keypoints_2d` containing the body part locations and detection confidence formatted as `x1,y1,c1,x2,y2,c2,...`. The coordinates `x` and `y` can be normalized to the range [0,1], [-1,1], [0, source size], [0, output size], etc., depending on the flag `keypoint_scale` (see flag for more information), while `c` is the confidence score in the range [0,1].
2. The arrays `face_keypoints_2d`, `hand_left_keypoints_2d`, and `hand_right_keypoints_2d`, analogous to `pose_keypoints_2d`.
3. The analogous 3-D arrays `body_keypoints_3d`, `face_keypoints_3d`, `hand_left_keypoints_2d`, and `hand_right_keypoints_2d` (if `--3d` is enabled, otherwise they will be empty). Instead of `x1,y1,c1,x2,y2,c2,...`, their format is `x1,y1,z1,c1,x2,y2,z2,c2,...`, where `c` is simply 1 or 0 depending on whether the 3-D reconstruction was successful or not.
4. The body part candidates before being assembled into people (if `--part_candidates` is enabled).
2. Part candidates (`--part_candidates`) are saved with the same scale than the final keypoints itself.
3. Fixed bug in keepTopNPeople.hpp (`--number_people_max`) that provoked core dumped if lots of values equal to the threshold.
4. Flir cameras: Cameras sorted by serial number. Video and images recorded from flir cameras were (and are) assigned the camera parameters based on serial number order, so it would fail if the cameras order was not the same than if sorted by serial number.
5. CPU version working in non-Nvidia Windows machines.
