OCRNet Recognition Issues in DeepStream 8.0

[kbesenic-vt]

Hi! I am porting the nvOCDR solution to DeepStream 8.0. I managed to port the solution to DeepStream 7.0 and it worked flawlessly. However, I am encountering significant issues with the OCRNet recognition stage on 8.0.

Observed Behavior:

• OCDNet functions correctly; bounding boxes are accurately localized.

• OCRNet produces scrambled or fragmented text readings, like it's reading from corrupted or shifted memory.

Technical Context:

• Environment: DeepStream 8.0 devel container (nvcr.io/nvidia/deepstream:8.0-gc-triton-devel).

• Pipeline: Using the run_demo.sh from the Dockerfile.

• TensorRT: Engines were successfully regenerated for TensorRT 10.x.

Can you tell me if there are some changes in DeepStream 8.0 that require explicit handling in the custom library compared to versions 7.0 and 6.2?