tensorrt, SSD, NMS with different width height feature maps

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Hello,
I have succesfully converted a custom tensorflow object detection ssd model to tensorrt 4.0.16.
I’m running an input with 1600x1600 and its output is as expected.
I want to change the input to 1920x1080 and have accordingly change config.py and the plugin code to support that.

Input = gs.create_node(
    "Input",
    op="Placeholder",
    dtype=tf.float32,
    # batch, channels, height, width
    shape=[1, 3, 1080, 1920])

PriorBox = gs.create_node(
    "PriorBox",
    numLayers=6,
    minScale=0.1,
    maxScale=0.2,
    aspectRatios=[1.0, 2.0, 0.5, 3.0, 0.33],
    layerVariances=[0.1, 0.1, 0.2, 0.2],
    featureMapShapesWidth=[240, 120, 60, 30, 15, 7],
    featureMapShapesHeight=[135, 68, 34, 17, 9, 4])

The serialized engine is correctly created and loaded.
The problem is that i keep getting saturated x0,x1 values (y0,y1 is correct) to 1.

The question is NMS been tested with different feature map sizes for width and height ? Is this a known bug ?

2

why my posts are hidden ?

3

I have the similar issues. Have you solved this problem?

4

maybe feature maps are not correct

5

hey, were you able to address this problem? Can you provide more details of the model you were using and the sample code you based the config.py file from?