TensorRT-7.2.3.4.Ubuntu-18.04.x86-64-gnu.cuda-11.1.cudnn8.1

# Object Detection With A TensorFlow SSD Network **Table Of Contents** - [Description](#description) - [How does this sample work?](#how-does-this-sample-work) * [Processing the input graph](#processing-the-input-graph) * [Preparing the data](#preparing-the-data) * [sampleUffSSD plugins](#sampleuffssd-plugins) * [Verifying the output](#verifying-the-output) * [TensorRT API layers and ops](#tensorrt-api-layers-and-ops) - [Prerequisites](#prerequisites) - [Running the sample](#running-the-sample) * [Sample `--help` options](#sample-help-options) - [Additional resources](#additional-resources) - [License](#license) - [Changelog](#changelog) - [Known issues](#known-issues) ## Description This sample, sampleUffSSD, preprocesses a TensorFlow SSD network, performs inference on the SSD network in TensorRT, using TensorRT plugins to speed up inference. This sample is based on the [SSD: Single Shot MultiBox Detector](https://arxiv.org/abs/1512.02325) paper. The SSD network performs the task of object detection and localization in a single forward pass of the network. The SSD network used in this sample is based on the TensorFlow implementation of SSD, which actually differs from the original paper, in that it has an inception_v2 backbone. For more information about the actual model, download [ssd_inception_v2_coco](http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2017_11_17.tar.gz). The TensorFlow SSD network was trained on the InceptionV2 architecture using the [MSCOCO dataset](http://cocodataset.org/#home) which has 91 classes (including the background class). The config details of the network can be found [here](https://github.com/tensorflow/models/blob/master/research/object_detection/samples/configs/ssd_inception_v2_coco.config). ## How does this sample work? The SSD network performs the task of object detection and localization in a single forward pass of the network. The TensorFlow SSD network was trained on the InceptionV2 architecture using the MSCOCO dataset. The sample makes use of TensorRT plugins to run the SSD network. To use these plugins, the TensorFlow graph needs to be preprocessed, and we use the GraphSurgeon utility to do this. The main components of this network are the Image Preprocessor, FeatureExtractor, BoxPredictor, GridAnchorGenerator and Postprocessor. **Image Preprocessor** The image preprocessor step of the graph is responsible for resizing the image. The image is resized to a 300x300x3 size tensor. This step also performs normalization of the image so all pixel values lie between the range [-1, 1]. **FeatureExtractor** The FeatureExtractor portion of the graph runs the InceptionV2 network on the preprocessed image. The feature maps generated are used by the anchor generation step to generate default bounding boxes for each feature map. In this network, the size of feature maps that are used for anchor generation are [(19x19), (10x10), (5x5), (3x3), (2x2), (1x1)]. **BoxPredictor** The BoxPredictor step takes in a high level feature map as input and produces a list of box encodings (x-y coordinates) and a list of class scores for each of these encodings per feature map. This information is passed to the postprocessor. **GridAnchorGenerator** The goal of this step is to generate a set of default bounding boxes (given the scale and aspect ratios mentioned in the config) for each feature map cell. This is implemented as a plugin layer in TensorRT called the `gridAnchorGenerator` plugin. The registered plugin name is `GridAnchor_TRT`. **Postprocessor** The postprocessor step performs the final steps to generate the network output. The bounding box data and confidence scores for all feature maps are fed to the step along with the pre-computed default bounding boxes (generated in the `GridAnchorGenerator` namespace). It then performs NMS (non-maximum suppression) which prunes away most of the bounding boxes based on a confidence threshold and IoU (Intersection over Union) overlap, thus storing only the top `N` boxes per class. This is implemented as a plugin layer in TensorRT called the NMS plugin. The registered plugin name is `NMS_TRT`. **Note:** This sample also implements another plugin called `FlattenConcat` which is used to flatten each input and then concatenate the results. This is applied to the location and confidence data before it is fed to the post processor step since the NMS plugin requires the data to be in this format. For details on how a plugin is implemented, see the implementation of `FlattenConcat` plugin and `FlattenConcatPluginCreator` in the `sampleUffSSD.cpp` file in the `tensorrt/samples/sampleUffSSD` directory. Specifically, this sample performs the following steps: - [Processing the input graph](#processing-the-input-graph) - [Preparing the data](#preparing-the-data) - [sampleUffSSD plugins](#sampleuffssd-plugins) - [Verifying the output](#verifying-the-output) ### Processing the input graph The TensorFlow SSD graph has some operations that are currently not supported in TensorRT. Using a preprocessor on the graph, we can combine multiple operations in the graph into a single custom operation which can be implemented as a plugin layer in TensorRT. Currently, the preprocessor provides the ability to stitch all nodes within a namespace into one custom node. To use the preprocessor, the `convert-to-uff` utility should be called with a `-p` flag and a config file. The config script should also include attributes for all custom plugins which will be embedded in the generated `.uff` file. Current sample script for SSD is located in `/usr/src/tensorrt/samples/sampleUffSSD/config.py`. Using the preprocessor on the graph, we were able to remove the `Preprocessor` namespace from the graph, stitch the `GridAnchorGenerator` namespace together to create the `GridAnchorGenerator` plugin, stitch the `postprocessor` namespace together to get the NMS plugin and mark the concat operations in the BoxPredictor as `FlattenConcat` plugins. The TensorFlow graph has some operations like `Assert` and `Identity` which can be removed for the inferencing. Operations like `Assert` are removed and leftover nodes (with no outputs once assert is deleted) are then recursively removed. `Identity` operations are deleted and the input is forwarded to all the connected outputs. Additional documentation on the graph preprocessor can be found in the [TensorRT API](https://docs.nvidia.com/deeplearning/sdk/tensorrt-api/python_api/graphsurgeon/graphsurgeon.html). ### Preparing the data The generated network has an input node called `Input`, and the output node is given the name `MarkOutput_0` by the UFF converter. These nodes are registered by the UFF Parser in the sample. ``` parser->registerInput("Input", DimsCHW(3, 300, 300), UffInputOrder::kNCHW); parser->registerOutput("MarkOutput_0"); ``` The input to the SSD network in this sample is 3 channel 300x300 images. In the sample, we normalize the image so the pixel values lie in the range [-1,1]. This is equivalent to the image preprocessing stage of the network. Since TensorRT does not depend on any computer vision libraries, the images are represented in binary `R`, `G`, and `B` values for each pixel. The format is Portable PixMap (PPM), which is a netpbm color image format. In this format, the `R`, `G`, and `B` values for each pixel are represented by a byte of integer (0-255) and they are stored together, pixel by pixel. There is a simple PPM reading function called `readPPMFile`. ### sampleUffSSD plugins Details about how to create TensorRT plugins can be found in [Extending TensorRT With Custom Layers](https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#extending). The `config.py` defined for the `convert-to-uff` command should have the custom layers mapped to the plugin names in TensorRT by modifying the `op` field. The names of the plugin parameters should also exactly match those expected