Deepstream issue with certain rtsp streams?

Cameras: Hikvision

Hi,
We are currently dealing with a scenario where deepstream is failing to detect objects when run on rtsp stream but the same stream when recorded and played via a file input source, the detection quality seems to be significantly better. And this is something observed with some specific cameras. Any idea around this issue or is it a known issue?

Which cameras? Can you identify the types and specs?

@Fiona.Chen,

The cameras info is as below,

Model and Make - HikVision Bullet cameras.

The camera information that I get from the hikvision portal, I’ve attached an image below, please take a look

We could not find anything out of the blue when compared to other cameras where detections looked fine. Thanks. Please let me know if you need any other information

What app is you using? If it is deepstream-app, can you send us the configuration files? Can you use filesink to record a piece of the video after inferrence and send us?

Hi beefshepherd,

Is this still an issue to support? Any update?

@kayccc @Fiona.Chen,
Sorry, was caught up with something else. Yes this issue still persists. And detections are an amiss on rtsp feeds but recorded videos seemed to be working fine. Here is a link of the detections recorded from DS,
video . Please let me know if any other information is required. Thanks.

What app is you using? If it is deepstream-app, can you send us the configuration files?

@Fiona.Chen, I am currently using a modified version of the app, the config file for the detector is as follows,
[property]
gpu-id=0
net-scale-factor=0.0039215697906911373
model-file=resnet10.caffemodel
proto-file=resnet10.prototxt
#model-engine-file=person_detector_nvidia.caffemodel_b1_int8.engine
labelfile-path=labels.txt
int8-calib-file=cal_trt.bin
batch-size=1
model-color-format=0
process-mode=1
network-mode=2
num-detected-classes=4
interval=3
gie-unique-id=1
output-blob-names=conv2d_bbox;conv2d_cov/Sigmoid

[class-attrs-0]
threshold=1.2
eps=0.2
group-threshold=1

[class-attrs-1]
threshold=1.2
eps=0.2
group-threshold=1

[class-attrs-2]
threshold=0.2
eps=0.2
group-threshold=1

[class-attrs-3]
threshold=1.2
eps=0.2
group-threshold=1,

What we need is deepstream-app configuration file but not nvinfer config file.
If there is only source change with the pipeline, you may need to check the camera captured video quality first. If the captured video is with poor quality (e.g. with too much noise), it will impact inference result.