Practical overview
When it fits
Use Dd3D when you need a code-based research workflow for monocular 3D detection, particularly for training or evaluating experiments on KITTI or nuScenes. It is suited to work in a local GPU and dataset environment rather than calling a hosted inference service.
Who it is for
- Computer-vision researchers studying monocular 3D object detection.
- Developers evaluating or adapting a PyTorch detection research workflow.
- Teams comparing monocular detection experiments with pseudo-LiDAR approaches.
Documented uses
- Train a monocular 3D object detection experiment using supported dataset configurations.
- Evaluate a pretrained model with an experiment configuration and local dataset.
- Visualize the data loader and check a training loop through the documented validation workflows.
- Configure research runs with modular Hydra settings and single-node multi-GPU training or evaluation.
Requirements & limits
The repository is public and licensed under MIT. The documented workflow uses a local Docker environment; optional AWS caching and Weights & Biases integration require associated accounts and configuration.
- Docker is recommended, and the README identifies Ubuntu 18.04 as the only tested operating system.
- The setup instructions distinguish CUDA 10.2 and CUDA 11.1 Dockerfiles and advise checking NVIDIA driver/CUDA compatibility.
- KITTI and nuScenes must be downloaded and organized under the configured dataset root.
- Pretrained depth-estimation weights are listed for DLA34, V2-99 and OmniML backbones.
- The training script supports single-node multi-GPU training and evaluation through MPI.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up the documented Docker environment
Clone the repository and build its recommended Docker environment for subsequent experiments.
Before you start
- A terminal and Git
- Docker and NVIDIA Docker support
- An NVIDIA driver and CUDA-compatible setup
- The README specifies Ubuntu 18.04 as the only tested operating system.
Steps
Clone the repository with the command below and enter the
dd3ddirectory.git clone https://github.com/TRI-ML/dd3d.gitBuild the recommended CUDA 10.2 image with the command below.
make docker-buildFor the documented CUDA 11.1 alternative, use the command below and check driver/CUDA compatibility when choosing the Dockerfile.
make docker-build DOCKERFILE=Dockerfile-cu111
Expected outcome
A Docker image prepared as the project environment; the README's subsequent commands are intended to run inside the container.
Guide limitations
- The README says the environment was tested only on Ubuntu 18.04.
- This guide covers setup only; it does not run training or evaluation.
Evaluate a pretrained model
Run the documented evaluation-only workflow using a selected experiment configuration and pretrained checkpoint.
Before you start
- The Dd3D environment set up according to the repository instructions
- A dataset arranged at the configured dataset root
- An experiment configuration and pretrained model checkpoint
- A compatible GPU setup; the README also gives a smaller single-GPU test batch option.
Steps
Choose an experiment configuration and a pretrained checkpoint.
Run
./scripts/train.py +experiments= EVAL_ONLY=True MODEL.CKPT=.For the README's smaller single-GPU batch example, add
TEST.IMS_PER_BATCH=4.
Expected outcome
The documented evaluation-only command runs the selected experiment configuration with the specified checkpoint.
Guide limitations
- The command uses placeholders that must be replaced with the chosen experiment and checkpoint path.
- The README does not describe this as a hosted inference endpoint.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- TRI-ML/dd3d · Official homepage title · retained in its source language