NVIDIA-ISAAC-ROS
Need help on testing hardware
We have been able to run Python version of Foundationpose released by NVLabs on an RTX3060 With 12GB Ram whereas when we try ROS version of foundationpose the GPU memory requirements suits up and ultimately the execution stops with below errors:
component_container_mt-1] 2024-06-21 14:47:16.877 ERROR gxf/std/block_memory_pool.cpp@77: Failure in cudaMalloc. cuda_error: cudaErrorMemoryAllocation, error_str: out of memory
[component_container_mt-1] 2024-06-21 14:47:16.877 ERROR gxf/std/entity_warden.cpp@437: Failed to initialize component 00157 (pool)
[component_container_mt-1] 2024-06-21 14:47:16.877 ERROR gxf/core/runtime.cpp@702: Could not initialize entity 'YNWEMQPEYV_inference' (E152): GXF_OUT_OF_MEMORY
[component_container_mt-1] 2024-06-21 14:47:16.877 ERROR gxf/std/program.cpp@283: Failed to activate entity 00152 named YNWEMQPEYV_inference: GXF_OUT_OF_MEMORY
[component_container_mt-1] 2024-06-21 14:47:16.877 ERROR gxf/std/program.cpp@285: Deactivating...
[component_container_mt-1] 2024-06-21 14:47:16.877 ERROR gxf/core/runtime.cpp@1452: Graph activation failed with error: GXF_OUT_OF_MEMORY
Is there a fix/workaround of this? Is there any reason for memory requirement of this package to go higher compare to original Foundationpose?
The requirement mentioned by you on the documentation is already satisfied. https://nvidia-isaac-ros.github.io/repositories_and_packages/isaac_ros_pose_estimation/index.html#supported-platforms
+1 We are facing the same issue. Were any of you able to resolve it?
We have built a custom version of FoundationPose on ROS2 with Conda environment that works even with 8GB of GPU:
https://github.com/ammar-n-abbas/FoundationPoseROS2
Thank you @ammar-n-abbas for sharing the repo! Just curious how much FPS were you able to achieve in best case end to end with SAM2?
Hi @yashwant-ies SAM2 is only used in the first frame to extract the masks. We achieved 10FPS with this model on an 8GB GeForce RTX 4070.