I’m encountering a consistent CUDA crash when calling retrieve_objects() in ZED SDK 5.0, using CUSTOM_BOX_OBJECTS
and resolution set to VGA
.
The crash occurs even without ingesting any custom detections — just initializing the object detection module with CUSTOM_BOX_OBJECTS
and retrieving sl.Objects()
causes a fatal error on the GPU.
This does not happen with higher resolutions (e.g. HD720
).
Only occurs with:
camera_resolution = sl.RESOLUTION.VGA
object_detection_model = sl.OBJECT_DETECTION_MODEL.CUSTOM_BOX_OBJECTS
depth_mode = sl.DEPTH_MODE.NEURAL
orNEURAL_LIGHT
(tested with both)- Works fine at
HD720
and above - Reproduced consistently with a minimal Python script using the ZED SDK 5.0
- Happens even without using
ingest_custom_mask_objects()
- Crash occurs on
grab()
Minimal reproducible Python script:
import pyzed.sl as sl
zed = sl.Camera()
init_params = sl.InitParameters()
init_params.camera_resolution = sl.RESOLUTION.VGA
init_params.depth_mode = sl.DEPTH_MODE.NEURAL # also tested with ULTRA
init_params.coordinate_units = sl.UNIT.METER
init_params.sdk_verbose = True
assert zed.open(init_params) == sl.ERROR_CODE.SUCCESS
assert zed.enable_positional_tracking() == sl.ERROR_CODE.SUCCESS
obj_params = sl.ObjectDetectionParameters()
obj_params.detection_model = sl.OBJECT_DETECTION_MODEL.CUSTOM_BOX_OBJECTS
assert zed.enable_object_detection(obj_params) == sl.ERROR_CODE.SUCCESS
runtime_params = sl.RuntimeParameters()
obj_runtime_params = sl.ObjectDetectionRuntimeParameters()
while True:
if zed.grab(runtime_params) == sl.ERROR_CODE.SUCCESS:
objects = sl.Objects()
zed.retrieve_objects(objects, obj_runtime_params)
System Info:
- GPU: NVIDIA RTX 4060 Ti
- Driver: 570.133.20
- OS: Ubuntu 24.04
- ZED SDK: 5.0.0 (Python API)
- Python: 3.9
- CUDA: 12.8