Parichay Kapoor

>src pad: 1 ~ 16 cropped other/tensor-flex according to crop info > or 1 original tensor + 15 tensor-flex How do you plan to handle a varying number of output...

For other/tensor-flex, using `GstMeta` to store the dynamic size information can be helpful to minimize the memory requirement. `GstParentBufferMeta` can be used to refer to the parent buffer holding the...

> You may need to update /debian/control and /debian/rules to properly control this. Yes, that's needed and will be done.

> Q: Could we build pytorch/caffe2 v1.6.0 with protobuf 3.12.3 for 16.04? Yes we can locally with python 3.6. However, it fails on launchpad. Python 3.5 is not supported anymore....

1. QoS - given multiple models, use the model (given a list of heaviest to lightest to choose from) based on performance being achieved and update the model being used...

Support openGL to eliminate CPU-GPU-CPU data copy. 1. In single shot API, this can directly take GPU data and output the same. 2. In pipeline API, Filter, source and sink...

There is a group of plugins in gstreamer based on opengl - https://gstreamer.freedesktop.org/documentation/opengl/index.html?gi-language=c

Support **auto-complete pipeline** feature to make the process of making pipelines easy. This takes input as incomplete pipelines, and gives completed pipelines as output, with missing elements inserted in pipeline,...

> In order to support this feature, we need to modify the gstreamer framework, don't we? We can provide this as a part of NNStreamer API to be used with...

Supporting valgrind without `--leak-check=full` seems feasible with some of the current test cases I tried (unittest_common, plugins, repo, sink and src_iio). However, running with valgrind exposes a deadlock issue inside...