duplicate with 299559 and 301792
1.resnet18_3d_rgb_hmdb5_32.etlt is a classification model, So there will be no bbox and no coordinates.
2.If you want to get the frame number, use the following patch
diff --git a/sources/apps/sample_apps/deepstream-3d-action-recognition/deepstream_3d_action_recognition.cpp b/sources/apps/sample_apps/deepstream-3d-action-recognition/deepstream_3d_action_recognition.cpp
index 9ef1c1e..03bd523 100644
--- a/sources/apps/sample_apps/deepstream-3d-action-recognition/deepstream_3d_action_recognition.cpp
+++ b/sources/apps/sample_apps/deepstream-3d-action-recognition/deepstream_3d_action_recognition.cpp
@@ -91,9 +91,9 @@ add_fps_display_meta(NvDsFrameMeta *frame, NvDsBatchMeta *batch_meta) {
nvds_add_display_meta_to_frame(frame, display_meta);
}
+static int frame = 0;
/* tiler_sink_pad_buffer_probe will extract metadata received on OSD sink pad
* and update params for drawing rectangle, object information etc. */
-
static GstPadProbeReturn
pgie_src_pad_buffer_probe(GstPad *pad, GstPadProbeInfo *info,
gpointer u_data)
@@ -103,6 +103,7 @@ pgie_src_pad_buffer_probe(GstPad *pad, GstPadProbeInfo *info,
NvDsMetaList *l_user_meta = NULL;
NvDsUserMeta *user_meta = NULL;
+ frame++;
for (l_user_meta = batch_meta->batch_user_meta_list; l_user_meta != NULL;
l_user_meta = l_user_meta->next)
{
@@ -121,6 +122,7 @@ pgie_src_pad_buffer_probe(GstPad *pad, GstPadProbeInfo *info,
}
for (auto &roi_meta : preprocess_batchmeta->roi_vector)
{
+ // printf("%.2f %.2f %.2f %.2f \n", roi_meta.roi.left, roi_meta.roi.top, roi_meta.roi.width, roi_meta.roi.height);
NvDsMetaList *l_user = NULL;
for (l_user = roi_meta.roi_user_meta_list; l_user != NULL;
l_user = l_user->next)
@@ -143,6 +145,9 @@ pgie_src_pad_buffer_probe(GstPad *pad, GstPadProbeInfo *info,
const gchar *label = "";
if (class_id < MAX_CLASS_LEN)
label = kActioClasseLabels[class_id];
+ if (!strncasecmp(label, "fall_floor", strlen("fall_floor"))) {
+ printf("fall_floor %d \n", frame);
+ }
LOG_DEBUG("output tensor result: cls_id: %d, scrore:%.3f, label: %s", class_id, max_prob, label);
}
}
Using the test video from this link, you can get the following results.
fall_floor 96
fall_floor 97
fall_floor 100
fall_floor 101
fall_floor 102
fall_floor 103
fall_floor 104
fall_floor 105
fall_floor 106
fall_floor 107
fall_floor 108
3.I use Jetpack 6.0 and Deepstream 7.0 on AGX Orin. For better performance, you can try setting AGX Orin to maxn model