How to have good quality resized frame in NVIDIA Jetson Nano B01 similar to Raspberry Pi 4B?

[jasonrichdarmawan]

My goal is to mimic $ libcamera-still --width 1280 --height 1280 -o pi.jpg which only available on Raspberry Pi 4B.

The problem is that the frame produced in NVIDIA Jetson Nano B01 is significantly poorer than the frame produced in Raspberry Pi 4B. This maybe because of gstreamer_pipeline as I can’t find good documentation regarding this.

My expected result is that the resized frame have good detail when zoomed by 700%. This is because the frame produced in Raspberry Pi 4B have good detail when zoomed by 700%.

The actual result is that the resized frame do not have any detail.

Frame produced in Raspberry Pi 4B

Frame produced in Jetson Nano B01

What I’ve tried is:

1. use this repository ArduCAM/Jetson_IMX519_Focus_Example (github.com)
2. Then, rewrite the repository to only support some features that I need.
3. Change the gstreamer_pipeline function

def gstreamer_pipeline(
    # Issue: the sensor format used by Raspberry Pi 4B and NVIDIA Jetson Nano B01 are different
    # in Raspberry Pi 4B, this command
    # $ libcamera-still --width 1280 --height 1280 --mode 1280:1280
    # uses sensor format 2328x1748.
    # However, v4l2-ctl --list-formats-ext do not have such format.
    sensor_id=0,
    capture_width=4656,
    capture_height=3496,
    # known issue:
    # if the frame_rate is set to the suggested frame rate by `$ v4l2-ctl --list-formats-ext``
    # the captured frame will have black pixels at the bottom of the frame.
    # the solution:
    # subtract the suggested frame rate by 1.
    frame_rate=10-1,
    display_width=1280,
    display_height=1280
):
    return (
        f"nvarguscamerasrc sensor-id={sensor_id} ! "
        f"video/x-raw(memory:NVMM),width=(int){capture_width},height=(int){capture_height},format=(string)NV12,framerate=(fraction){frame_rate}/1 ! "
        
        # comment if prefer the crop strategy.
        f"nvvidconv flip-method=0 ! "
        
        # uncomment if prefer the crop strategy.
        # f"nvvidconv flip-method=0 "
        # f"left={int(capture_width/2 - display_width/2)} "
        # f"right={int(capture_width/2 + display_width/2)} "
        # f"top={int(capture_height/2 - display_height/2)} "
        # f"bottom={int(capture_height/2 + display_height/2)} ! "
        # f"video/x-raw,width=(int){display_width},height=(int){display_height},format=(string)BGRx ! "
        
        # comment if prefer the crop strategy.
        f"video/x-raw,width=(int){display_width},height=(int){display_height},format=(string)BGRx ! "
        
        f"videoconvert ! "
        f"video/x-raw,format=(string)BGR ! "
        f"appsink"
    )

[DaneLLL]

Hi,
You may try to set this property in nvvidconv plugin:

  interpolation-method: Set interpolation methods
                        flags: readable, writable, controllable
                        Enum "GstInterpolationMethod" Default: 0, "Nearest"
                           (0): Nearest          - Nearest
                           (1): Bilinear         - Bilinear
                           (2): 5-Tap            - 5-Tap
                           (3): 10-Tap           - 10-Tap
                           (4): Smart            - Smart
                           (5): Nicest           - Nicest

These are all modes supported in hardware converter. Please try and see if you can get desired result.

It is possible that certain hardware blocks are not as good as other embedded platforms, but our GPU blocks shall perform better than others.