| 1 |
/* |
| 2 |
* GStreamer gstreamer-yolotensordecoder |
| 3 |
* Copyright (C) 2024 Collabora Ltd. |
| 4 |
* Authors: Daniel Morin <[email protected]> |
| 5 |
* Vineet Suryan <[email protected]> |
| 6 |
* Santosh Mahto <[email protected]> |
| 7 |
* |
| 8 |
* gstyolotensordecoder.c |
| 9 |
* |
| 10 |
* This library is free software; you can redistribute it and/or |
| 11 |
* modify it under the terms of the GNU Library General Public |
| 12 |
* License as published by the Free Software Foundation; either |
| 13 |
* version 2 of the License, or (at your option) any later version. |
| 14 |
* |
| 15 |
* This library is distributed in the hope that it will be useful, |
| 16 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 17 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 18 |
* Library General Public License for more details. |
| 19 |
* |
| 20 |
* You should have received a copy of the GNU Library General Public |
| 21 |
* License along with this library; if not, write to the |
| 22 |
* Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, |
| 23 |
* Boston, MA 02110-1301, USA. |
| 24 |
*/ |
| 25 |
|
| 26 |
/** |
| 27 |
* SECTION:element-yolotensordec |
| 28 |
* @short_description: Decode tensors from YOLO detection models |
| 29 |
* |
| 30 |
* This element can parse per-buffer inference tensors meta data generated |
| 31 |
* by an upstream inference element |
| 32 |
* |
| 33 |
* |
| 34 |
* ## Example launch command: |
| 35 |
* |
| 36 |
* Test image file, model file and labels file can be found here : |
| 37 |
* https://gitlab.collabora.com/gstreamer/onnx-models |
| 38 |
* |
| 39 |
* gst-launch-1.0 -v v4l2src \ |
| 40 |
* ! videoconvertscale qos=false ! video/x-raw, pixel-aspect-ratio=1/1 \ |
| 41 |
* ! onnxinference model-file=yolov8s.onnx \ |
| 42 |
* ! yolov8tensordec class-confidence-threshold=0.8 iou-threshold=0.3 \ |
| 43 |
* max-detections=100 label-file=labels/COCO_classes.txt \ |
| 44 |
* ! objectdetectionoverlay ! glimagesink sink=gtkglsink |
| 45 |
* |
| 46 |
* The original repository of the Yolo is located at |
| 47 |
* https://github.com/ultralytics/ultralytics. |
| 48 |
* For easy experimentation, the models based on Yolo architecture in Onnx |
| 49 |
* format can be found at https://col.la/gstonnxmodels . This model already |
| 50 |
* has tensors name embedded matching default values of tensors-detections-name |
| 51 |
* and tensors-logits-name properties. It's also possible to embed tensor-ids |
| 52 |
* into any model based on Yolo architecture to allow this tensor-decoder |
| 53 |
* to decode tensors.This process is described in the Readme of |
| 54 |
* repository: https://col.la/gstonnxmodels" |
| 55 |
* |
| 56 |
* Since: 1.28 |
| 57 |
*/ |
| 58 |
|
| 59 |
#ifdef HAVE_CONFI_H |
| 60 |
#include "config.h" |
| 61 |
#endif |
| 62 |
|
| 63 |
#include "gstyolotensordecoder.h" |
| 64 |
|
| 65 |
#include <gst/analytics/analytics.h> |
| 66 |
#include <gst/analytics/gstanalytics_image_util.h> |
| 67 |
#include <gio/gio.h> |
| 68 |
|
| 69 |
#include <math.h> |
| 70 |
|
| 71 |
#define YOLO_DETECTION_MASK "yolo-v8-out" |
| 72 |
GQuark YOLO_DETECTION_MASK_ID; |
| 73 |
|
| 74 |
/** |
| 75 |
* GstYoloTensorDecoder: |
| 76 |
* |
| 77 |
* A tensor decoder for YOLO v8-v11 models. |
| 78 |
* |
| 79 |
* Since: 1.28 |
| 80 |
*/ |
| 81 |
|
| 82 |
GST_DEBUG_CATEGORY_STATIC (yolo_tensor_decoder_debug); |
| 83 |
#define GST_CAT_DEFAULT yolo_tensor_decoder_debug |
| 84 |
|
| 85 |
GST_ELEMENT_REGISTER_DEFINE (yolo_tensor_decoder, "yolov8tensordec", |
| 86 |
GST_RANK_PRIMARY, GST_TYPE_YOLO_TENSOR_DECODER); |
| 87 |
|
| 88 |
/* GstYoloTensorDecoder properties, see properties description in |
| 89 |
* gst_yolo_tensor_decoder_class_init for more details. */ |
| 90 |
enum |
| 91 |
{ |
| 92 |
PROP_0, |
| 93 |
PROP_BOX_CONFI_THRESH, |
| 94 |
PROP_CLS_CONFI_THRESH, |
| 95 |
PROP_IOU_THRESH, |
| 96 |
PROP_MAX_DETECTION, |
| 97 |
PROP_LABEL_FILE |
| 98 |
}; |
| 99 |
|
| 100 |
/* Specify the range of confidence level in tensor output*/ |
| 101 |
typedef struct _ConfidenceRange |
| 102 |
{ |
| 103 |
gsize start; /* Start index of confidence level */ |
| 104 |
gsize end; /* End index of confidence level */ |
| 105 |
gsize step; /* Step size of next confidence level index */ |
| 106 |
} ConfidenceRange; |
| 107 |
|
| 108 |
/* Default properties value */ |
| 109 |
static const gfloat DEFAULT_BOX_CONFI_THRESH = 0.4f; |
| 110 |
static const gfloat DEFAULT_CLS_CONFI_THRESH = 0.4f; |
| 111 |
static const gfloat DEFAULT_IOU_THRESH = 0.7f; |
| 112 |
static const gsize DEFAULT_MAX_DETECTION = 100; |
| 113 |
|
| 114 |
/* Global variable storing class for OD. Generally OD has class |
| 115 |
* and we need to provide one but this class is just a placeholder.*/ |
| 116 |
GQuark OOI_CLASS_ID; |
| 117 |
|
| 118 |
/* GStreamer element srcpad template. Template of a srcpad that can receive |
| 119 |
* any raw video. */ |
| 120 |
static GstStaticPadTemplate gst_yolo_tensor_decoder_src_template = |
| 121 |
GST_STATIC_PAD_TEMPLATE ("src", |
| 122 |
GST_PAD_SRC, |
| 123 |
GST_PAD_ALWAYS, |
| 124 |
GST_STATIC_CAPS ("video/x-raw")); |
| 125 |
|
| 126 |
/* *INDENT-OFF* */ |
| 127 |
static GstStaticPadTemplate gst_yolo_tensor_decoder_sink_template = |
| 128 |
GST_STATIC_PAD_TEMPLATE ("sink", |
| 129 |
GST_PAD_SINK, |
| 130 |
GST_PAD_ALWAYS, |
| 131 |
GST_STATIC_CAPS ( |
| 132 |
"video/x-raw," |
| 133 |
"tensors=(structure)[" |
| 134 |
"tensorgroups," |
| 135 |
"yolo-v8-out=(/uniquelist){" |
| 136 |
"(GstCaps)[" |
| 137 |
"tensor/strided," |
| 138 |
"tensor-id=(string)yolo-v8-out," |
| 139 |
"dims=<(int)1,(int)[1,max],(int)[1,max]>," |
| 140 |
"dims-order=(string)col-major," |
| 141 |
"type=(string)float32" |
| 142 |
"]" |
| 143 |
"}" |
| 144 |
"]" |
| 145 |
)); |
| 146 |
/* *INDENT-ON* */ |
| 147 |
|
| 148 |
/* GstYoloTensorDecoder Prototypes */ |
| 149 |
static gboolean gst_yolo_tensor_decoder_set_caps (GstBaseTransform * trans, |
| 150 |
GstCaps * incaps, GstCaps * outcaps); |
| 151 |
static void gst_yolo_tensor_decoder_set_property (GObject * object, |
| 152 |
guint prop_id, const GValue * value, GParamSpec * pspec); |
| 153 |
static void gst_yolo_tensor_decoder_get_property (GObject * object, |
| 154 |
guint prop_id, GValue * value, GParamSpec * pspec); |
| 155 |
static GstFlowReturn gst_yolo_tensor_decoder_transform_ip (GstBaseTransform * |
| 156 |
trans, GstBuffer * buf); |
| 157 |
static void gst_yolo_tensor_decoder_finalize (GObject * object); |
| 158 |
|
| 159 |
static void gst_yolo_tensor_decoder_object_found (GstYoloTensorDecoder * self, |
| 160 |
GstAnalyticsRelationMeta * rmeta, BBox * bb, gfloat confidence, |
| 161 |
GQuark class_quark, const gfloat * candidate_masks, gsize offset, |
| 162 |
guint count); |
| 163 |
|
| 164 |
G_DEFINE_TYPE (GstYoloTensorDecoder, gst_yolo_tensor_decoder, |
| 165 |
GST_TYPE_BASE_TRANSFORM); |
| 166 |
|
| 167 |
static GArray * |
| 168 |
read_labels (const char *labels_file) |
| 169 |
{ |
| 170 |
GArray *array; |
| 171 |
GFile *file = g_file_new_for_path (labels_file); |
| 172 |
GFileInputStream *file_stream; |
| 173 |
GDataInputStream *data_stream; |
| 174 |
GError *error = NULL; |
| 175 |
gchar *line; |
| 176 |
|
| 177 |
file_stream = g_file_read (file, NULL, &error); |
| 178 |
g_object_unref (file); |
| 179 |
if (!file_stream) { |
| 180 |
GST_WARNING ("Could not open file %s: %s\n", labels_file, error->message); |
| 181 |
g_clear_error (&error); |
| 182 |
return NULL; |
| 183 |
} |
| 184 |
|
| 185 |
data_stream = g_data_input_stream_new (G_INPUT_STREAM (file_stream)); |
| 186 |
g_object_unref (file_stream); |
| 187 |
|
| 188 |
array = g_array_new (FALSE, FALSE, sizeof (GQuark)); |
| 189 |
|
| 190 |
while ((line = g_data_input_stream_read_line (data_stream, NULL, NULL, |
| 191 |
&error))) { |
| 192 |
GQuark label = g_quark_from_string (line); |
| 193 |
g_array_append_val (array, label); |
| 194 |
g_free (line); |
| 195 |
} |
| 196 |
|
| 197 |
g_object_unref (data_stream); |
| 198 |
|
| 199 |
if (error) { |
| 200 |
GST_WARNING ("Could not open file %s: %s", labels_file, error->message); |
| 201 |
g_array_free (array, TRUE); |
| 202 |
g_clear_error (&error); |
| 203 |
return NULL; |
| 204 |
} |
| 205 |
|
| 206 |
if (array->len == 0) { |
| 207 |
g_array_free (array, TRUE); |
| 208 |
return NULL; |
| 209 |
} |
| 210 |
|
| 211 |
return array; |
| 212 |
} |
| 213 |
|
| 214 |
static gboolean |
| 215 |
gst_yolo_tensor_decoder_set_caps (GstBaseTransform * trans, GstCaps * incaps, |
| 216 |
GstCaps * outcaps) |
| 217 |
{ |
| 218 |
GstYoloTensorDecoder *self = GST_YOLO_TENSOR_DECODER (trans); |
| 219 |
|
| 220 |
if (!gst_video_info_from_caps (&self->video_info, incaps)) { |
| 221 |
GST_ERROR_OBJECT (self, "Failed to parse caps"); |
| 222 |
return FALSE; |
| 223 |
} |
| 224 |
|
| 225 |
if (gst_base_transform_is_passthrough (trans)) { |
| 226 |
GST_ERROR_OBJECT (self, "Failed. Can't handle passthrough"); |
| 227 |
return FALSE; |
| 228 |
} |
| 229 |
|
| 230 |
return TRUE; |
| 231 |
} |
| 232 |
|
| 233 |
static void |
| 234 |
gst_yolo_tensor_decoder_set_property (GObject * object, guint prop_id, |
| 235 |
const GValue * value, GParamSpec * pspec) |
| 236 |
{ |
| 237 |
GstYoloTensorDecoder *self = GST_YOLO_TENSOR_DECODER (object); |
| 238 |
const gchar *filename; |
| 239 |
|
| 240 |
switch (prop_id) { |
| 241 |
case PROP_BOX_CONFI_THRESH: |
| 242 |
self->box_confi_thresh = g_value_get_float (value); |
| 243 |
break; |
| 244 |
case PROP_CLS_CONFI_THRESH: |
| 245 |
self->cls_confi_thresh = g_value_get_float (value); |
| 246 |
break; |
| 247 |
case PROP_IOU_THRESH: |
| 248 |
self->iou_thresh = g_value_get_float (value); |
| 249 |
break; |
| 250 |
case PROP_MAX_DETECTION: |
| 251 |
self->max_detection = g_value_get_uint (value); |
| 252 |
break; |
| 253 |
case PROP_LABEL_FILE: |
| 254 |
{ |
| 255 |
GArray *labels; |
| 256 |
|
| 257 |
filename = g_value_get_string (value); |
| 258 |
labels = read_labels (filename); |
| 259 |
|
| 260 |
if (labels) { |
| 261 |
g_free (self->label_file); |
| 262 |
self->label_file = g_strdup (filename); |
| 263 |
g_clear_pointer (&self->labels, g_array_unref); |
| 264 |
self->labels = labels; |
| 265 |
} else { |
| 266 |
GST_WARNING_OBJECT (self, "Label file '%s' not found!", filename); |
| 267 |
} |
| 268 |
break; |
| 269 |
} |
| 270 |
default: |
| 271 |
G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec); |
| 272 |
break; |
| 273 |
} |
| 274 |
} |
| 275 |
|
| 276 |
static void |
| 277 |
gst_yolo_tensor_decoder_get_property (GObject * object, guint prop_id, |
| 278 |
GValue * value, GParamSpec * pspec) |
| 279 |
{ |
| 280 |
GstYoloTensorDecoder *self = GST_YOLO_TENSOR_DECODER (object); |
| 281 |
|
| 282 |
switch (prop_id) { |
| 283 |
case PROP_BOX_CONFI_THRESH: |
| 284 |
g_value_set_float (value, self->box_confi_thresh); |
| 285 |
break; |
| 286 |
case PROP_CLS_CONFI_THRESH: |
| 287 |
g_value_set_float (value, self->cls_confi_thresh); |
| 288 |
break; |
| 289 |
case PROP_IOU_THRESH: |
| 290 |
g_value_set_float (value, self->iou_thresh); |
| 291 |
break; |
| 292 |
case PROP_MAX_DETECTION: |
| 293 |
g_value_set_uint (value, self->max_detection); |
| 294 |
break; |
| 295 |
case PROP_LABEL_FILE: |
| 296 |
g_value_set_string (value, self->label_file); |
| 297 |
break; |
| 298 |
default: |
| 299 |
G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec); |
| 300 |
break; |
| 301 |
} |
| 302 |
} |
| 303 |
|
| 304 |
static void |
| 305 |
gst_yolo_tensor_decoder_class_init (GstYoloTensorDecoderClass * klass) |
| 306 |
{ |
| 307 |
GObjectClass *gobject_class = (GObjectClass *) klass; |
| 308 |
GstElementClass *element_class = (GstElementClass *) klass; |
| 309 |
GstBaseTransformClass *basetransform_class = (GstBaseTransformClass *) klass; |
| 310 |
|
| 311 |
/* Define GstYoloTensorDecoder debug category. */ |
| 312 |
GST_DEBUG_CATEGORY_INIT (yolo_tensor_decoder_debug, |
| 313 |
"yolov8tensordec", 0, "Tensor decoder for Yolo detection models"); |
| 314 |
|
| 315 |
YOLO_DETECTION_MASK_ID = g_quark_from_static_string (YOLO_DETECTION_MASK); |
| 316 |
|
| 317 |
/* Set GObject vmethod to get and set property */ |
| 318 |
gobject_class->set_property = gst_yolo_tensor_decoder_set_property; |
| 319 |
gobject_class->get_property = gst_yolo_tensor_decoder_get_property; |
| 320 |
|
| 321 |
g_object_class_install_property (G_OBJECT_CLASS (klass), |
| 322 |
PROP_BOX_CONFI_THRESH, |
| 323 |
g_param_spec_float ("box-confidence-threshold", |
| 324 |
"Box location confidence threshold", |
| 325 |
"Boxes with a location confidence level inferior to this threshold " |
| 326 |
"will be excluded", |
| 327 |
0.0, 1.0, DEFAULT_BOX_CONFI_THRESH, |
| 328 |
(GParamFlags) (G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS))); |
| 329 |
|
| 330 |
g_object_class_install_property (G_OBJECT_CLASS (klass), |
| 331 |
PROP_CLS_CONFI_THRESH, |
| 332 |
g_param_spec_float ("class-confidence-threshold", |
| 333 |
"Class confidence threshold", |
| 334 |
"Classes with a confidence level inferior to this threshold " |
| 335 |
"will be excluded", |
| 336 |
0.0, 1.0, DEFAULT_CLS_CONFI_THRESH, |
| 337 |
(GParamFlags) (G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS))); |
| 338 |
|
| 339 |
g_object_class_install_property (G_OBJECT_CLASS (klass), |
| 340 |
PROP_IOU_THRESH, |
| 341 |
g_param_spec_float ("iou-threshold", |
| 342 |
"Maximum IOU threshold", |
| 343 |
"Maximum intersection-over-union between bounding boxes to " |
| 344 |
"consider them distinct.", |
| 345 |
0.0, 1.0, DEFAULT_IOU_THRESH, |
| 346 |
(GParamFlags) (G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS))); |
| 347 |
|
| 348 |
g_object_class_install_property (G_OBJECT_CLASS (klass), |
| 349 |
PROP_MAX_DETECTION, |
| 350 |
g_param_spec_uint ("max-detections", |
| 351 |
"Maximum object/masks detections.", |
| 352 |
"Maximum object/masks detections.", |
| 353 |
1, G_MAXUINT, DEFAULT_MAX_DETECTION, |
| 354 |
(GParamFlags) (G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS))); |
| 355 |
|
| 356 |
g_object_class_install_property (G_OBJECT_CLASS (klass), PROP_LABEL_FILE, |
| 357 |
g_param_spec_string ("label-file", |
| 358 |
"Label file", "Label file", NULL, (GParamFlags) |
| 359 |
(G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS))); |
| 360 |
|
| 361 |
|
| 362 |
/* Element description. */ |
| 363 |
gst_element_class_set_static_metadata (element_class, |
| 364 |
"YOLO v8-11 object detection tensor decoder", "Tensordecoder/Video", |
| 365 |
"Decode tensors output from the inference of YOLO Object Detection or FastSAM model (Detection)" |
| 366 |
"on video frames. This works on YOLO version 8 and later(v11), and FastSAM models.", |
| 367 |
"Daniel Morin <[email protected]>, Santosh Mahto <[email protected]>"); |
| 368 |
|
| 369 |
/* Add pads to element base on pad template defined earlier */ |
| 370 |
gst_element_class_add_pad_template (element_class, |
| 371 |
gst_static_pad_template_get (&gst_yolo_tensor_decoder_src_template)); |
| 372 |
gst_element_class_add_pad_template (element_class, |
| 373 |
gst_static_pad_template_get (&gst_yolo_tensor_decoder_sink_template)); |
| 374 |
|
| 375 |
/* Set GstBaseTransform vmethod transform_ip. This methode is called |
| 376 |
* by the srcpad when it receive buffer. ip stand for in-place meaning the |
| 377 |
* buffer remain unchanged by the element. Tensor-decoder only monitor |
| 378 |
* buffer it receive for a meta attach to the buffer that is a GstTensorMeta |
| 379 |
* and has a tensor-id can be handled by GstYoloTensorDecoder. */ |
| 380 |
basetransform_class->transform_ip = |
| 381 |
GST_DEBUG_FUNCPTR (gst_yolo_tensor_decoder_transform_ip); |
| 382 |
|
| 383 |
/* Set GstBaseTransform set_caps vmethod. This will be called once the |
| 384 |
* capability negotiation has been completed. We will be able to extract |
| 385 |
* resolution from this callback. */ |
| 386 |
basetransform_class->set_caps = |
| 387 |
GST_DEBUG_FUNCPTR (gst_yolo_tensor_decoder_set_caps); |
| 388 |
|
| 389 |
gobject_class->finalize = gst_yolo_tensor_decoder_finalize; |
| 390 |
|
| 391 |
klass->object_found = gst_yolo_tensor_decoder_object_found; |
| 392 |
|
| 393 |
/* Calculate the class id placeholder (also a quark) that will be set |
| 394 |
* as label if object if labels are not provided via label-file. */ |
| 395 |
OOI_CLASS_ID = g_quark_from_static_string ("Yolo-None"); |
| 396 |
} |
| 397 |
|
| 398 |
struct Candidate |
| 399 |
{ |
| 400 |
const float *candidate; |
| 401 |
const float max_confidence; |
| 402 |
const guint max_class_offset; |
| 403 |
}; |
| 404 |
|
| 405 |
static void |
| 406 |
gst_yolo_tensor_decoder_init (GstYoloTensorDecoder * self) |
| 407 |
{ |
| 408 |
/* GstYoloTensorDecoder instance initialization */ |
| 409 |
self->box_confi_thresh = DEFAULT_BOX_CONFI_THRESH; |
| 410 |
self->cls_confi_thresh = DEFAULT_CLS_CONFI_THRESH; |
| 411 |
self->iou_thresh = DEFAULT_IOU_THRESH; |
| 412 |
self->max_detection = DEFAULT_MAX_DETECTION; |
| 413 |
|
| 414 |
self->sel_candidates = g_array_new (FALSE, FALSE, sizeof (struct Candidate)); |
| 415 |
self->selected = g_ptr_array_new (); |
| 416 |
|
| 417 |
gst_base_transform_set_passthrough (GST_BASE_TRANSFORM (self), FALSE); |
| 418 |
} |
| 419 |
|
| 420 |
static const GstTensor * |
| 421 |
gst_yolo_tensor_decoder_get_tensor (GstYoloTensorDecoder * |
| 422 |
self, GstBuffer * buf) |
| 423 |
{ |
| 424 |
GstMeta *meta = NULL; |
| 425 |
gpointer iter_state = NULL; |
| 426 |
|
| 427 |
if (!gst_buffer_get_meta (buf, GST_TENSOR_META_API_TYPE)) { |
| 428 |
GST_DEBUG_OBJECT (self, |
| 429 |
"missing tensor meta from buffer %" GST_PTR_FORMAT, buf); |
| 430 |
return NULL; |
| 431 |
} |
| 432 |
|
| 433 |
while ((meta = gst_buffer_iterate_meta_filtered (buf, &iter_state, |
| 434 |
GST_TENSOR_META_API_TYPE))) { |
| 435 |
GstTensorMeta *tensor_meta = (GstTensorMeta *) meta; |
| 436 |
const GstTensor *tensor; |
| 437 |
const gsize YOLO_DETECTIONS_TENSOR_N_DIMS = 3; |
| 438 |
static const gsize dims[3] = { 1, G_MAXSIZE, G_MAXSIZE }; |
| 439 |
|
| 440 |
tensor = gst_tensor_meta_get_typed_tensor (tensor_meta, |
| 441 |
YOLO_DETECTION_MASK_ID, GST_TENSOR_DATA_TYPE_FLOAT32, |
| 442 |
GST_TENSOR_DIM_ORDER_COL_MAJOR, YOLO_DETECTIONS_TENSOR_N_DIMS, dims); |
| 443 |
|
| 444 |
if (tensor) { |
| 445 |
if (tensor->dims[1] < 5) { |
| 446 |
GST_WARNING_OBJECT (self, "Ignore tensor because dims[1] is %zu < 5", |
| 447 |
tensor->dims[1]); |
| 448 |
continue; |
| 449 |
} |
| 450 |
|
| 451 |
return tensor; |
| 452 |
} |
| 453 |
} |
| 454 |
|
| 455 |
return NULL; |
| 456 |
} |
| 457 |
|
| 458 |
|
| 459 |
static GstFlowReturn |
| 460 |
gst_yolo_tensor_decoder_transform_ip (GstBaseTransform * trans, GstBuffer * buf) |
| 461 |
{ |
| 462 |
GstYoloTensorDecoder *self = GST_YOLO_TENSOR_DECODER (trans); |
| 463 |
GstAnalyticsRelationMeta *rmeta; |
| 464 |
const GstTensor *detections_tensor; |
| 465 |
|
| 466 |
detections_tensor = gst_yolo_tensor_decoder_get_tensor (self, buf); |
| 467 |
if (detections_tensor == NULL) { |
| 468 |
GST_WARNING_OBJECT (self, "Couldn't find mask tensor, skipping"); |
| 469 |
return GST_FLOW_OK; |
| 470 |
} |
| 471 |
|
| 472 |
/* Retrieve or attach an analytics-relation-meta to the buffer. |
| 473 |
* Analytics-relation-meta are container that can reveive multiple |
| 474 |
* analytics-meta, like OD and Segmentation. The following call will only |
| 475 |
* retrieve an analytics-relation-meta if it exist or create one if it |
| 476 |
* does not exist. */ |
| 477 |
rmeta = gst_buffer_add_analytics_relation_meta (buf); |
| 478 |
if (rmeta == NULL) { |
| 479 |
GST_ELEMENT_ERROR (trans, STREAM, FAILED, (NULL), |
| 480 |
("Analytics Relation meta allocation failed")); |
| 481 |
return GST_FLOW_ERROR; |
| 482 |
} |
| 483 |
|
| 484 |
if (!gst_yolo_tensor_decoder_decode_f32 (self, rmeta, detections_tensor, 0)) |
| 485 |
return GST_FLOW_ERROR; |
| 486 |
|
| 487 |
return GST_FLOW_OK; |
| 488 |
} |
| 489 |
|
| 490 |
static void |
| 491 |
gst_yolo_tensor_decoder_finalize (GObject * object) |
| 492 |
{ |
| 493 |
GstYoloTensorDecoder *self = GST_YOLO_TENSOR_DECODER (object); |
| 494 |
|
| 495 |
g_clear_pointer (&self->sel_candidates, g_array_unref); |
| 496 |
g_clear_pointer (&self->selected, g_ptr_array_unref); |
| 497 |
|
| 498 |
g_free (self->label_file); |
| 499 |
g_clear_pointer (&self->labels, g_array_unref); |
| 500 |
|
| 501 |
G_OBJECT_CLASS (gst_yolo_tensor_decoder_parent_class)->finalize (object); |
| 502 |
} |
| 503 |
|
| 504 |
/* Extract bounding box from tensor data */ |
| 505 |
static void |
| 506 |
gst_yolo_tensor_decoder_convert_bbox (const gfloat * candidate, |
| 507 |
const gsize * offset, BBox * bbox) |
| 508 |
{ |
| 509 |
gfloat w = *(candidate + offset[2]); |
| 510 |
gfloat h = *(candidate + offset[3]); |
| 511 |
bbox->x = *(candidate + offset[0]) - (w / 2); |
| 512 |
bbox->y = *(candidate + offset[1]) - (h / 2); |
| 513 |
bbox->w = w + 0.5; |
| 514 |
bbox->h = h + 0.5; |
| 515 |
} |
| 516 |
|
| 517 |
/* Calculate iou between boundingbox of candidate c1 and c2 |
| 518 |
*/ |
| 519 |
static gfloat |
| 520 |
gst_yolo_tensor_decoder_iou (const gfloat * c1, const gfloat * c2, |
| 521 |
const gsize * offset, BBox * bb1, BBox * bb2) |
| 522 |
{ |
| 523 |
gst_yolo_tensor_decoder_convert_bbox (c1, offset, bb1); |
| 524 |
gst_yolo_tensor_decoder_convert_bbox (c2, offset, bb2); |
| 525 |
return gst_analytics_image_util_iou_int (bb1->x, bb1->y, bb1->w, bb1->h, |
| 526 |
bb2->x, bb2->y, bb2->w, bb2->h); |
| 527 |
} |
| 528 |
|
| 529 |
/* Utility function to find maxmum confidence value across classes |
| 530 |
* specified by range. |
| 531 |
*/ |
| 532 |
static gfloat |
| 533 |
gst_yolo_tensor_decoder_find_max_class_confidence (const gfloat * c, |
| 534 |
const ConfidenceRange * c_range, gsize * max_class_ofs) |
| 535 |
{ |
| 536 |
gfloat max_val = 0.0; |
| 537 |
for (gsize i = c_range->start; i <= c_range->end; i += c_range->step) { |
| 538 |
if (*(c + i) > max_val) { |
| 539 |
max_val = *(c + i); |
| 540 |
*max_class_ofs = i; |
| 541 |
} |
| 542 |
} |
| 543 |
return max_val; |
| 544 |
} |
| 545 |
|
| 546 |
/* Compare c1 and c2 |
| 547 |
* Utility function for sorting candiates based on the a field identified |
| 548 |
* by offset. |
| 549 |
*/ |
| 550 |
static gint |
| 551 |
gst_yolo_tensor_decoder_sort_candidates (gconstpointer p1, gconstpointer p2) |
| 552 |
{ |
| 553 |
const struct Candidate *c1 = p1; |
| 554 |
const struct Candidate *c2 = p2; |
| 555 |
|
| 556 |
if (c1->max_confidence < c2->max_confidence) |
| 557 |
return 1; |
| 558 |
else if (c1->max_confidence > c2->max_confidence) |
| 559 |
return -1; |
| 560 |
else |
| 561 |
return 0; |
| 562 |
} |
| 563 |
|
| 564 |
static gboolean |
| 565 |
gst_yolo_tensor_decoder_decode_valid_bb (GstYoloTensorDecoder * self, |
| 566 |
gfloat x, gfloat y, gfloat w, gfloat h) |
| 567 |
{ |
| 568 |
GstYoloTensorDecoder *parent = GST_YOLO_TENSOR_DECODER (self); |
| 569 |
|
| 570 |
if (x > (GST_VIDEO_INFO_WIDTH (&parent->video_info))) |
| 571 |
return FALSE; |
| 572 |
if (y > (GST_VIDEO_INFO_HEIGHT (&parent->video_info))) |
| 573 |
return FALSE; |
| 574 |
if (x < -(gfloat) (GST_VIDEO_INFO_WIDTH (&parent->video_info) / 2.0)) |
| 575 |
return FALSE; |
| 576 |
if (y < -(gfloat) (GST_VIDEO_INFO_HEIGHT (&parent->video_info) / 2.0)) |
| 577 |
return FALSE; |
| 578 |
if (w <= 0) |
| 579 |
return FALSE; |
| 580 |
if (h <= 0) |
| 581 |
return FALSE; |
| 582 |
if (w > (GST_VIDEO_INFO_WIDTH (&parent->video_info))) |
| 583 |
return FALSE; |
| 584 |
if (h > (GST_VIDEO_INFO_HEIGHT (&parent->video_info))) |
| 585 |
return FALSE; |
| 586 |
|
| 587 |
return TRUE; |
| 588 |
} |
| 589 |
|
| 590 |
gboolean |
| 591 |
gst_yolo_tensor_decoder_decode_f32 (GstYoloTensorDecoder * self, |
| 592 |
GstAnalyticsRelationMeta * rmeta, const GstTensor * detections_tensor, |
| 593 |
guint num_masks) |
| 594 |
{ |
| 595 |
GstMapInfo map_info_detections; |
| 596 |
gfloat iou; |
| 597 |
gboolean rv, keep; |
| 598 |
gsize offset, x_offset, y_offset, w_offset, h_offset, offsets[4]; |
| 599 |
BBox bb1, bb2; |
| 600 |
ConfidenceRange c_range; |
| 601 |
gsize max_class_offset = 0, class_index; |
| 602 |
GQuark class_quark = OOI_CLASS_ID; |
| 603 |
gboolean ret = TRUE; |
| 604 |
gsize i; |
| 605 |
|
| 606 |
/* Retrieve memory at index 0 and map it in READWRITE mode */ |
| 607 |
rv = gst_buffer_map (detections_tensor->data, &map_info_detections, |
| 608 |
GST_MAP_READ); |
| 609 |
if (rv == FALSE) { |
| 610 |
GST_ELEMENT_ERROR (self, STREAM, FAILED, (NULL), |
| 611 |
("Could not map tensor buffer %" GST_PTR_FORMAT, |
| 612 |
detections_tensor->data)); |
| 613 |
return FALSE; |
| 614 |
} |
| 615 |
|
| 616 |
GST_LOG_OBJECT (self, "Mask Tensor shape dims %zu", |
| 617 |
detections_tensor->num_dims); |
| 618 |
|
| 619 |
/* Trace detections tensor dimensions */ |
| 620 |
if (gst_debug_category_get_threshold (GST_CAT_DEFAULT) >= GST_LEVEL_TRACE) { |
| 621 |
for (gsize i = 0; i < detections_tensor->num_dims; i++) { |
| 622 |
GST_TRACE_OBJECT (self, "Detections Tensor dim %zu: %zu", i, |
| 623 |
detections_tensor->dims[i]); |
| 624 |
} |
| 625 |
} |
| 626 |
|
| 627 |
/* Number of candidates can be large, reset the arrays */ |
| 628 |
g_array_set_size (self->sel_candidates, 0); |
| 629 |
g_ptr_array_set_size (self->selected, 0); |
| 630 |
|
| 631 |
/* detections_tensor->dims[2] contain the number of candidates. Let's call the |
| 632 |
* number of candidates C. We store this value in offset as we use it |
| 633 |
* calculate the offset of candidate fields. The variable #data_detections above point |
| 634 |
* at the detections tensor data, but candidates data is organize like a plane. |
| 635 |
* Candidates bbox X coord fields from 0 to C start at the begining of the |
| 636 |
* tensor data and are continguous in memory, followed by all candidates |
| 637 |
* field Y, followed by field W, ... followed by field class confidence level, |
| 638 |
* ..., followed by all candidates mask0, ..., followed by all candidates |
| 639 |
* mask31. Bellow we pre-calculate each field offset relative to the |
| 640 |
* candidate pointer (pointer to field X), which will allow us to easily |
| 641 |
* access each candiates field. |
| 642 |
* */ |
| 643 |
offset = detections_tensor->dims[2]; |
| 644 |
x_offset = 0; |
| 645 |
y_offset = offset; |
| 646 |
w_offset = 2 * offset; |
| 647 |
h_offset = 3 * offset; |
| 648 |
/* first index that contain confidence level */ |
| 649 |
c_range.start = 4 * offset; |
| 650 |
/* Last index that contains confidence level */ |
| 651 |
c_range.end = (detections_tensor->dims[1] - num_masks - 1) * offset; |
| 652 |
/* Step between class confidence level */ |
| 653 |
c_range.step = offset; |
| 654 |
offsets[0] = x_offset; |
| 655 |
offsets[1] = y_offset; |
| 656 |
offsets[2] = w_offset; |
| 657 |
offsets[3] = h_offset; |
| 658 |
|
| 659 |
#define BB_X(candidate) candidate[x_offset] |
| 660 |
#define BB_Y(candidate) candidate[y_offset] |
| 661 |
#define BB_W(candidate) candidate[w_offset] |
| 662 |
#define BB_H(candidate) candidate[h_offset] |
| 663 |
|
| 664 |
for (gsize c_idx = 0; c_idx < detections_tensor->dims[2]; c_idx++) { |
| 665 |
float *candidate = (float *) map_info_detections.data; |
| 666 |
|
| 667 |
candidate += c_idx; |
| 668 |
|
| 669 |
/* Yolo have multiple class, so maximum confidence level across all class is used |
| 670 |
* to evaluate the relevance of the candidate. Here we filter candidates |
| 671 |
* based on their class confidence level.*/ |
| 672 |
gfloat max_confidence = |
| 673 |
gst_yolo_tensor_decoder_find_max_class_confidence (candidate, &c_range, |
| 674 |
&max_class_offset); |
| 675 |
if (max_confidence > self->cls_confi_thresh |
| 676 |
&& gst_yolo_tensor_decoder_decode_valid_bb (self, |
| 677 |
BB_X (candidate), BB_Y (candidate), BB_W (candidate), |
| 678 |
BB_H (candidate))) { |
| 679 |
|
| 680 |
struct Candidate c = { |
| 681 |
candidate, |
| 682 |
max_confidence, |
| 683 |
max_class_offset, |
| 684 |
}; |
| 685 |
g_array_append_val (self->sel_candidates, c); |
| 686 |
|
| 687 |
GST_TRACE_OBJECT (self, "%zu: x,y=(%f;%f) w,h=(%f;%f), s=%f c=%f", |
| 688 |
c_idx, candidate[x_offset], candidate[y_offset], |
| 689 |
candidate[w_offset], candidate[h_offset], |
| 690 |
candidate[w_offset] * candidate[h_offset], max_confidence); |
| 691 |
} |
| 692 |
|
| 693 |
/* Pointer arithmetic, going to the next candidate. This is the candidate |
| 694 |
* pointer that is now incremented to the next candidate which is also |
| 695 |
* the field X of the next candidate.*/ |
| 696 |
candidate += 1; |
| 697 |
} |
| 698 |
|
| 699 |
GST_LOG_OBJECT (self, "Before NMS selected candidates count: %u", |
| 700 |
self->sel_candidates->len); |
| 701 |
|
| 702 |
/* We sort the remaining candidates because, in the next selection phase we |
| 703 |
* have a maximum and we want to make sure that considered only the candidates |
| 704 |
* with the highest class confidence level before potentially reaching the |
| 705 |
* maximum.*/ |
| 706 |
g_array_sort (self->sel_candidates, gst_yolo_tensor_decoder_sort_candidates); |
| 707 |
|
| 708 |
if (gst_debug_category_get_threshold (GST_CAT_DEFAULT) >= GST_LEVEL_TRACE) { |
| 709 |
for (i = 0; i < self->sel_candidates->len; i++) { |
| 710 |
struct Candidate *c = &g_array_index (self->sel_candidates, |
| 711 |
struct Candidate, i); |
| 712 |
GST_TRACE_OBJECT (self, |
| 713 |
"Sorted: %zu: x,y=(%f;%f) w,h=(%f;%f), s=%f c=%f", |
| 714 |
i, |
| 715 |
c->candidate[x_offset], |
| 716 |
c->candidate[y_offset], |
| 717 |
c->candidate[w_offset], |
| 718 |
c->candidate[h_offset], |
| 719 |
c->candidate[w_offset] * c->candidate[h_offset], c->max_confidence); |
| 720 |
} |
| 721 |
} |
| 722 |
|
| 723 |
/* Algorithm in part inspired by OpenCV NMSBoxes */ |
| 724 |
for (i = 0; i < self->sel_candidates->len; i++) { |
| 725 |
const struct Candidate *c = &g_array_index (self->sel_candidates, |
| 726 |
struct Candidate, i); |
| 727 |
keep = TRUE; |
| 728 |
|
| 729 |
/* We only want to a NMS using IoU between candidates we've decided to |
| 730 |
* keep and the new one we considering to keep. selected array contain |
| 731 |
* the candidates we decided to keep and candidates[c] is the candidate |
| 732 |
* we're considering to keep or reject */ |
| 733 |
for (gsize s = 0; s < self->selected->len && keep; s++) { |
| 734 |
const float *candidate2 = g_ptr_array_index (self->selected, s); |
| 735 |
iou = gst_yolo_tensor_decoder_iou (c->candidate, candidate2, |
| 736 |
offsets, &bb1, &bb2); |
| 737 |
keep = (iou <= self->iou_thresh); |
| 738 |
} |
| 739 |
|
| 740 |
if (keep) { |
| 741 |
if (self->selected->len == 0) { |
| 742 |
/* The first bounding-box always get in as there's no others bbox |
| 743 |
* to filter on based on IoU */ |
| 744 |
gst_yolo_tensor_decoder_convert_bbox (c->candidate, offsets, &bb1); |
| 745 |
} |
| 746 |
|
| 747 |
g_ptr_array_add (self->selected, (gpointer) c->candidate); |
| 748 |
|
| 749 |
if (self->labels) { |
| 750 |
class_index = (c->max_class_offset - c_range.start) / c_range.step; |
| 751 |
|
| 752 |
if (class_index < self->labels->len) |
| 753 |
class_quark = g_array_index (self->labels, GQuark, class_index); |
| 754 |
} |
| 755 |
|
| 756 |
const gfloat *candidate_masks = NULL; |
| 757 |
if (num_masks) { |
| 758 |
/* detections weight will be stored in last `num_masks` |
| 759 |
* row of detections_tensor, so mask offset |
| 760 |
* will start at the end of the detections_tensor minus |
| 761 |
* `num_masks` |
| 762 |
*/ |
| 763 |
candidate_masks = c->candidate + |
| 764 |
((detections_tensor->dims[1] - num_masks) * offset); |
| 765 |
|
| 766 |
if (candidate_masks + num_masks + offset > |
| 767 |
(gfloat *) (map_info_detections.data + map_info_detections.size)) { |
| 768 |
GST_ELEMENT_ERROR (self, STREAM, FAILED, (NULL), |
| 769 |
("Mask tensor data size %zu is smaller than expected (%zu)", |
| 770 |
(candidate_masks - (gfloat *) map_info_detections.data) + |
| 771 |
offset + num_masks, map_info_detections.size)); |
| 772 |
ret = FALSE; |
| 773 |
break; |
| 774 |
} |
| 775 |
} |
| 776 |
|
| 777 |
GST_YOLO_TENSOR_DECODER_GET_CLASS (self)->object_found (self, rmeta, &bb1, |
| 778 |
c->max_confidence, class_quark, candidate_masks, offset, |
| 779 |
self->selected->len); |
| 780 |
|
| 781 |
/* If the maximum number of candidate selected is reached exit the |
| 782 |
* selection process. */ |
| 783 |
if (self->selected->len >= self->max_detection) { |
| 784 |
break; |
| 785 |
} |
| 786 |
} |
| 787 |
} |
| 788 |
|
| 789 |
GST_LOG_OBJECT (self, "After NMS selected count: %u", self->selected->len); |
| 790 |
|
| 791 |
/* We unmap the memory */ |
| 792 |
gst_buffer_unmap (detections_tensor->data, &map_info_detections); |
| 793 |
|
| 794 |
return ret; |
| 795 |
} |
| 796 |
|
| 797 |
static void |
| 798 |
gst_yolo_tensor_decoder_object_found (GstYoloTensorDecoder * self, |
| 799 |
GstAnalyticsRelationMeta * rmeta, BBox * bb, gfloat confidence, |
| 800 |
GQuark class_quark, const gfloat * candidate_masks, gsize offset, |
| 801 |
guint count) |
| 802 |
{ |
| 803 |
gst_analytics_relation_meta_add_od_mtd (rmeta, class_quark, bb->x, bb->y, |
| 804 |
bb->w, bb->h, confidence, NULL); |
| 805 |
} |