Pose Estimation from RGB Images of Highly Symmetric Objects using a Novel Multi-Pose Loss and Differential Rendering

Overview

We propose a novel multi-pose loss function to train a neural network for 6D pose estimation, using synthetic data and evaluating it on real images. Our loss is inspired by the VSD (Visible Surface Discrepancy) metric and relies on a differentiable renderer and CAD models. This novel multi-pose approach produces multiple weighted pose estimates to avoid getting stuck in local minima. Our method resolves pose ambiguities without using predefined symmetries. It is trained only on synthetic data. We test on real-world RGB images from the T-LESS dataset, containing highly symmetric objects common in industrial settings.

Setup

1. Install docker and setup nvidia runtime environment.

   • Install nvidia-docker2 (see https://gist.github.com/Brainiarc7/a8ab5f89494d053003454efc3be2d2ef)

   • Add "nvidia" as default runtime to build Docker image with CUDA support. Required for pytorch3d. (see https://stackoverflow.com/questions/59691207/docker-build-with-nvidia-runtime)

2. Build docker image(s)

• Clone this repo:

git clone [email protected]:shbe-aau/multi-pose-estimation.git

• Run bash build.sh in dockerfiles/pytorch3d to build the Docker image.
• (Optional) Run bash build.sh in dockerfiles/aae to build the Docker image for running the AAE (Augmented Autoencoder) - this is only needed for evaluation of the AAE approach.

Tested working on Ubuntu 16.04 LTS with Docker 18.09.7 and NVIDIA docker 2.3.0, and on Ubuntu 16.04 LTS with Docker 19.03.4 and NVIDIA docker 2.3.0.

A pre built image can be found at: https://hub.docker.com/r/shbe/pytorch3d_multi_pose

Prepare models and data

1. Download background images by executing the script download-voc2012.sh in multi-pose/data/VOC2012. We use the VOC2012 devkit for background images, they can also be found at http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#devkit

2. (Optional) Use the default for the T-LESS dataset already provided in multi-pose/data/cad-files or place your own CAD files in this folder.

3. (Optional) Prepare the encoder weights. You have the following options:

a. Use the default pretrained encoder weights already provided in 'multi-pose/data/encoder/obj1-18/encoder.npy'

b. Get the pretrained encoder as provided from the AAE repo.

c. Train your own autoencoder as specified in the AAE repo.

4. (Optional - only necessary for option 3b or 3c) Convert the trained encoder from Tensorflow to Pytorch format.
   • Instructions will be added later.

Train

1. Spin up the Docker container and start training:

   bash dockerfiles/pytorch3d/run-gpu0.sh
   cd shared-folder/multi-pose
   python train.py experiment_template.cfg
   

   Parameters for the training are set in the .cfg file found in 'multi-pose/experiments' folder. You can create new configurations files for each training session.

   The output of the training is saved at the path specified by the 'OUTPUT_PATH' parameter in the .cfg file. The default is './output/test' for 'experiment_template.cfg'. (NOTE: Remember to specify different output paths for different sessions of training or you may overwrite previous training sessions!)

Evaluating

1. Download the T-LESS 'BOP'19/20 test images' and 'Object models' from https://bop.felk.cvut.cz/datasets/ by running the script: download-tless.sh in multi-pose/data/tless/

2. Run the evaluation script:

   bash multi-pose/scripts/run-eval.sh \
   OBJ_ID \
   APPROACH_NAME \
   TRAINED_MODEL \
   DATA_SPLIT \
   DATASET
   

   Example - evaluating the first epoch from the 'experiment_template.cfg' run:

   bash multi-pose/scripts/run-eval.sh \
   10 \
   "experiment-template-epoch0" \
   "multi-pose/output/test/models/model-epoch0.pt" \
   "test" \
   "tless"
   

   NOTE: Do not use '_' for the APPROACH_NAME as it messes with the evaluation scripts.

3. The output of the evaluation can be found in multi-pose/data/tless/output/test-primesense/OBJ_ID/eval/

Visualize loss landscape

To visualize a loss landscape you can use the same config file as for training. Parameters under [Sampling] set number of points on a sphere to try, and which of those poses to treat as groundtruth (by index).

Determine poses and losses by running plot_loss_landscape.py in the docker container, and produce and view the final plots by running show_loss_landscape.py. You might want to run the second outside the docker to interact with the plots.