NVIDIA Object Detection Toolkit (ODTK)

Fast and accurate single stage object detection with end-to-end GPU optimization.

Description

ODTK is a single shot object detector with various backbones and detection heads. This allows performance/accuracy trade-offs.

It is optimized for end-to-end GPU processing using:

• The PyTorch deep learning framework with ONNX support
• NVIDIA Apex for mixed precision and distributed training
• NVIDIA DALI for optimized data pre-processing
• NVIDIA TensorRT for high-performance inference
• NVIDIA DeepStream for optimized real-time video streams support

Rotated bounding box detections

This repo now supports rotated bounding box detections. See rotated detections training and rotated detections inference documents for more information on how to use the --rotated-bbox command.

Bounding box annotations are described by [x, y, w, h, theta].

Performance

The detection pipeline allows the user to select a specific backbone depending on the latency-accuracy trade-off preferred.

ODTK RetinaNet model accuracy and inference latency & FPS (frames per seconds) for COCO 2017 (train/val) after full training schedule. Inference results include bounding boxes post-processing for a batch size of 1. Inference measured at --resize 800 using --with-dali on a FP16 TensorRT engine.

Installation

For best performance, use the latest PyTorch NGC docker container. Clone this repository, build and run your own image:

git clone https://github.com/nvidia/retinanet-examples
docker build -t odtk:latest retinanet-examples/
docker run --gpus all --rm --ipc=host -it odtk:latest

Usage

Training, inference, evaluation and model export can be done through the odtk utility. For more details, including a list of parameters, please refer to the TRAINING and INFERENCE documentation.

Training

Train a detection model on COCO 2017 from pre-trained backbone:

odtk train retinanet_rn50fpn.pth --backbone ResNet50FPN \
    --images /coco/images/train2017/ --annotations /coco/annotations/instances_train2017.json \
    --val-images /coco/images/val2017/ --val-annotations /coco/annotations/instances_val2017.json

Fine Tuning

Fine-tune a pre-trained model on your dataset. In the example below we use Pascal VOC with JSON annotations:

odtk train model_mydataset.pth --backbone ResNet50FPN \
    --fine-tune retinanet_rn50fpn.pth \
    --classes 20 --iters 10000 --val-iters 1000 --lr 0.0005 \
    --resize 512 --jitter 480 640 --images /voc/JPEGImages/ \
    --annotations /voc/pascal_train2012.json --val-annotations /voc/pascal_val2012.json

Note: the shorter side of the input images will be resized to resize as long as the longer side doesn't get larger than max-size. During training, the images will be randomly randomly resized to a new size within the jitter range.

Inference

Evaluate your detection model on COCO 2017:

odtk infer retinanet_rn50fpn.pth --images /coco/images/val2017/ --annotations /coco/annotations/instances_val2017.json

Run inference on your dataset:

odtk infer retinanet_rn50fpn.pth --images /dataset/val --output detections.json

Optimized Inference with TensorRT

For faster inference, export the detection model to an optimized FP16 TensorRT engine:

odtk export model.pth engine.plan

Evaluate the model with TensorRT backend on COCO 2017:

odtk infer engine.plan --images /coco/images/val2017/ --annotations /coco/annotations/instances_val2017.json

INT8 Inference with TensorRT

For even faster inference, do INT8 calibration to create an optimized INT8 TensorRT engine:

odtk export model.pth engine.plan --int8 --calibration-images /coco/images/val2017/

This will create an INT8CalibrationTable file that can be used to create INT8 TensorRT engines for the same model later on without needing to do calibration.

Or create an optimized INT8 TensorRT engine using a cached calibration table:

odtk export model.pth engine.plan --int8 --calibration-table /path/to/INT8CalibrationTable

Datasets

RetinaNet supports annotations in the COCO JSON format. When converting the annotations from your own dataset into JSON, the following entries are required:

{
    "images": [{
        "id" : int,
        "file_name" : str
    }],
    "annotations": [{
        "id" : int,
        "image_id" : int, 
        "category_id" : int,
        "bbox" : [x, y, w, h]   # all floats
        "area": float           # w * h. Required for validation scores
        "iscrowd": 0            # Required for validation scores
    }],
    "categories": [{
        "id" : int
    ]}
}

If using the --rotated-bbox flag for rotated detections, add an additional float theta to the annotations. To get validation scores you also need to fill the segmentation section.

        "bbox" : [x, y, w, h, theta]    # all floats, where theta is measured in radians anti-clockwise from the x-axis.
        "segmentation" : [[x1, y1, x2, y2, x3, y3, x4, y4]]
                                        # Required for validation scores.

Disclaimer

This is a research project, not an official NVIDIA product.

Jetpack compatibility

This branch uses TensorRT 7. If you are training and inferring models using PyTorch, or are creating TensorRT engines on Tesla GPUs (eg V100, T4), then you should use this branch.

If you wish to deploy your model to a Jetson device (eg - Jetson AGX Xavier) running Jetpack version 4.3, then you should use the 19.10 branch of this repo.

References

• Focal Loss for Dense Object Detection. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, Piotr Dollár. ICCV, 2017.
• Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, Kaiming He. June 2017.
• Feature Pyramid Networks for Object Detection. Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, Serge Belongie. CVPR, 2017.
• Deep Residual Learning for Image Recognition. Kaiming He, Xiangyu Zhang, Shaoqing Renm Jian Sun. CVPR, 2016.