Practical overview
When it fits
Use ODTK when you need a self-hosted training and inference toolkit for single-shot object detection and can supply labeled images in the documented annotation format. It is especially relevant when you want to tune backbone choices or use NVIDIA GPU-oriented training and inference components.
Who it is for
- Computer-vision practitioners training object detectors on labeled image datasets.
- Teams fine-tuning pretrained RetinaNet models or comparing backbone accuracy and latency trade-offs.
- Developers preparing TensorRT inference workflows for image detection or documented real-time video-stream support.
Documented uses
- Train a detector on COCO-format images and annotations.
- Fine-tune a pretrained detector on a custom dataset converted to COCO JSON.
- Run inference and save detections as JSON.
- Export a trained model to a TensorRT engine, including an INT8 calibration workflow.
- Train and infer with rotated bounding-box annotations.
Requirements & limits
The project is a public GitHub repository under the BSD-3-Clause license. The documented workflow is self-hosted and includes Docker commands with GPU access.
- Dataset annotations need to follow COCO JSON structure; rotated detections add an angle field.
- Backbone selection is presented as a way to trade detection accuracy against latency.
- A GPU-oriented Docker workflow is documented; TensorRT export and inference depend on having the relevant NVIDIA software and hardware environment.
- Published performance figures refer to specific model, dataset, resize, precision and hardware configurations.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Fine-tune a detector on a COCO-JSON dataset
Use a pretrained model and local images with COCO-format annotations to train a dataset-specific detector.
Before you start
- A built ODTK Docker image and a GPU-enabled Docker runtime, as shown in the installation instructions.
- A pretrained model file and training images.
- Training annotations in COCO JSON format; validation annotations are used in the documented example.
Steps
Clone the repository, build the Docker image and run it with GPU access using the README installation commands.
Prepare dataset image paths and COCO-format training and validation annotations.
Run the documented
odtk trainfine-tuning command with the pretrained model, backbone, class count, image paths and annotation paths.
Expected outcome
The documented workflow trains a fine-tuned detector on the supplied dataset.
Guide limitations
- The example fine-tuning command uses Pascal VOC data converted to JSON and specifies 20 classes; adapt those parameters to the dataset.
- COCO-format annotations require image, annotation and category records; validation scoring has additional annotation requirements.
Export a model for TensorRT inference
Export a trained detector as a TensorRT engine and run inference on images.
Before you start
- A trained model file and an image directory.
- The ODTK utility and TensorRT support in the installed environment.
Steps
Export the model with
odtk export model.pth engine.plan.Run inference with
odtk infer engine.plan --images /dataset/val --output detections.json.
Expected outcome
The documented commands create a TensorRT engine and write image detections to a JSON file.
Guide limitations
- The README presents the export as an optimized FP16 TensorRT engine; hardware and environment details affect practical performance.
- INT8 export requires calibration images or a cached calibration table.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 5 November 2026
- Original website title
- NVIDIA/retinanet-examples · Official homepage title · retained in its source language