Unable to suppress nvinfer object labels (text_params) in pyservicemaker probe — nvdsosd renders labels even when bbox is zeroed

AI & Data Science Computer Vision & Image Processing
,
1

Environment:

  • Jetson Orin Nano Super 8GB
  • JetPack 6.2.2
  • DeepStream 7.1
  • TensorRT 10.3
  • Python API: pyservicemaker (pybind11-based)
  • Pipeline: nvurisrcbin → nvstreammux → nvinfer → nvvideoconvert → capsfilter(RGBA) → nvdsosd → nv3dsink

What we’re doing:

We have an object detection pipeline (3 classes). We define rectangular lane ROIs on the frame. In a BatchMetadataOperator probe attached after nvinfer, we:

  1. Draw lane overlays on the OSD using display_meta.add_text() and display_meta.add_rect():
    def _draw_overlays(self, batch_meta, frame_meta, det_labels):
    display_meta = batch_meta.acquire_display_meta()

    for lane in self._lanes:
    rect = osd.Rect()
    rect.left = float(lane.x1)
    rect.top = float(lane.y1)
    rect.width = float(lane.x2 - lane.x1)
    rect.height = float(lane.y2 - lane.y1)
    rect.border_width = 3
    rect.border_color = color
    display_meta.add_rect(rect)

     text = osd.Text()                                                                                                                                          
 text.display_text = f"L{lane.id}".encode("ascii")                                                                                                          
 text.x_offset = int(lane.x1 + 5)                                                                                                                           
 text.y_offset = int(lane.y1 + 5)                                                                                                                           
 text.font.name = osd.FontFamily.Serif                                                                                                                      
 text.font.size = 14                                                                                                                                        
 text.font.color = osd.Color(1.0, 1.0, 1.0, 1.0)                                                                                                            
 display_meta.add_text(text)

    frame_meta.append(display_meta)

  2. Suppress bounding boxes outside lanes by zeroing rect_params — this works perfectly, nvdsosd does not render the bbox:
    for obj in frame_meta.object_items:
    r = obj.rect_params
    x1, y1 = float(r.left), float(r.top)
    w, h = float(r.width), float(r.height)
    cx, cy = x1 + w * 0.5, y1 + h * 0.5

    if not inside_any_lane(cx, cy):

    This successfully hides the bounding box

    r.width = 0
    r.height = 0
    r.border_width = 0
    continue

The problem:

Zeroing rect_params hides the bounding box, but nvdsosd still renders the text label (populated by nvinfer from labelfile-path) for the suppressed object. The
label floats outside the lane boundaries with no associated bbox.

What we’ve tried (all failed):

  1. obj.text_params.display_text = “” — Crashes with:
    TypeError: Unable to convert function return value to a Python type! The signature was
    (arg0: pyservicemaker._pydeepstream.ObjectMetadata) → _NvOSD_TextParams
  2. The text_params property getter in pyservicemaker’s pybind11 binding cannot convert _NvOSD_TextParams to a Python type.
  3. obj.label = “” — Does not affect nvdsosd rendering. nvdsosd reads from text_params.display_text, not the label property.
  4. Moving bbox off-screen (r.left = -10000.0, r.top = -10000.0) — nvdsosd appears to clamp negative coordinates to 0, so labels render at the top-left of the frame
    instead.
  5. Setting display-text=0 on nvdsosd element — This suppresses ALL text globally, including our custom display_meta.add_text() lane labels. We confirmed from the
    nvdsosd source (gstnvdsosd.c line 583) that draw_text gates both object metadata text AND display metadata text.

What we need:

A way to either:

  • Clear/suppress text_params.display_text on individual NvDsObjectMeta objects from a pyservicemaker probe
  • Or selectively disable nvinfer-generated object labels while keeping custom display_meta text rendering active

Potential workaround we’re considering:

Removing labelfile-path from the nvinfer config so nvinfer never populates text_params.display_text, then manually drawing labels only for in-lane detections via
display_meta.add_text(). However, we’d like to confirm this is the correct approach, or if there’s a proper way to modify text_params via pyservicemaker.

Pipeline setup (pyservicemaker):
p.add(“nvdsosd”, “osd”)

Probe attached to nvinfer output:

p.attach(“infer”, Probe(“my-probe”, probe))

nvinfer config (relevant lines):
labelfile-path=../../models/my_model/labels.txt
num-detected-classes=3
custom-lib-path=../../lib/my_parser.so
parse-bbox-func-name=MyCustomParser

Any guidance appreciated. Thanks!