SSD feature maps of different dimensions

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

Does the current SSDAnchorGenerator plugin support assymentric (different width height) feature maps ?
I’ve been experimenting with different sizes and only symmetric ones give correct results.

#2

To further explain,
I’ve created my own prior box plugin parser

PriorBox = gs.create_node(
    "PriorBox",
    numLayers=6,
    minScale=0.1,
    maxScale=0.9,
    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])

When the width matches the height feature map (for the correct sized frozen_inference_graph) i get the correct results, when they dont i the result x0 and x1 saturated to 1.

If this is an open bug with the SSDAnchorGenerator it would be great to know so i don’t spend any more time here.

#3

I have noticed the exact same problem with the plugin, if I use the same width and height as input dimensions everything works as expected, but as soon as I try other resolutions (e.g. 1920x1080) I get strange/wrong coordinate outputs. If someone could shed some light on this would be great.

#4

To be more specific i created a custom parser following the ssd example to parse assymetric feature maps so i can input resolutions of 1920x1080. The feature maps sizes are correct and verified by tensorboard and the engine is created.

PriorBox = gs.create_node(
    "PriorBox",
    numLayers=6,
    minScale=0.1,
    maxScale=0.9,
    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])

However the results whenever i have assymetric maps like above they are always wrong.

#5

The API seems to be able to handle different width/height

//!
    //! \brief The Anchor Generator plugin layer generates the prior boxes of designated sizes and aspect ratios across all dimensions @f$ (H \times W) @f$.
    //! GridAnchorParameters defines a set of parameters for creating the plugin layer for all feature maps.
    //! It contains:
    //! \param minScale Scale of anchors corresponding to finest resolution.
    //! \param maxScale Scale of anchors corresponding to coarsest resolution.
    //! \param aspectRatios List of aspect ratios to place on each grid point.
    //! \param numAspectRatios Number of elements in aspectRatios.
    //! \param H Height of feature map to generate anchors for.
    //! \param W Width of feature map to generate anchors for.
    //! \param variance Variance for adjusting the prior boxes.
    //!
    struct GridAnchorParameters
    {
        float minSize, maxSize;
        float *aspectRatios;
        int numAspectRatios, H, W;
        float variance[4];
    };

the code examples only use the same H,W for feature maps, I’m wondering if this is a bug width SSDAnchorGeneratorPlugin ?