Deep Object Pose Estimation (DOPE) - Inference

This directory contains a simple example of inference for DOPE.

Setup

If you haven't already, install the dependencies listed in requirements.txt in the root of the repo:

pip install -r ../requirements.txt

Running Inference

The inference.py script will take a trained model to run inference. In order to run, the following 3 arguments are needed:

--weights: path to the trained model weights. Can either point to a single .pth file or a folder containing multiple .pth files. If this path points to a folder with multiple .pth files, the script will individually load and run inference for all of the weights.
``--data: path to the data that will be used as input to run inference on. The script **recursively** loads all data that end with extensions specified in the --exts` flag.
--object: name of the class to run detections on. This name must be defined under dimensions in the config file passed to --config.

Below is an example of running inference:

python inference.py --weights ../weights --data ../sample_data --object cracker

Configuration Files

Depending on the images you want to run inference on, you may need to redefine the configuration values in camera_info.yaml and config_pose.yaml. You can either define a new configuration file and specify it with --config and --camera or update camera_info.yaml and config_pose.yaml.

Before running inference, it is important to make sure that:

The projection_matrix field is set properly in camera_info.yaml (or the file you specified for --camera). The projection_matrix field should be a 3x4 matrix of the form:

[fx,   0,  cx,  0,
  0,  fy,  cy,  0,
  0,   0,   1,  0]

The dimensions and class_ids fields have been specified for the object you wish to detect in config_pose.yaml (or the file you specified for --config).

Running Inference with Multiple Weights at Once

The inference script can run inference on multiple weights if the path specified in --weights points to a folder containing multiple .pth files. This feature is useful for fast evaluation of multiple weights to find the epoch that performs the best. While, generally, later epochs tend to perform better than earlier ones, this is not always the case. For more information on how to quantitatively evaluate the performance of a trained model, refer to the /evaluate subdirectory.