Skip to content

Latest commit

 

History

History
101 lines (59 loc) · 6.2 KB

README.md

File metadata and controls

101 lines (59 loc) · 6.2 KB
Raw

Computing Metrics

This document outlines how to compute the metrics we report in our paper, split into trajectory metrics and mapping metrics. We use a combination of external libraries and custom scripts to handle this.

The process for this is a bit convoluted. Feel free to open issues or email me at sethgi [at] umich [dot] edu with questions.

Trajectory Metrics

The process for computing trajectory metrics is - at a high level - as follows:

  1. Run prepare_results.py to put files into a useful directory structure
  2. Run analyze.sh to compute the metrics
  3. Run summarize_results.py to extract the information from the run and print to CSV or Tex.

Preparing Results

After running an ablation study using a combination of the --overrides and --num_repeats options, you will end up with several configurations (dicated by --overrides) and several trials per configuration (dictated by --num_repeats). All of that will be housed in a top-level experiment directory. The prepare_results.py script is responsible for taking information out of the top-level directory and moving it to a location that is useful for analysis. Run python3 prepare_results.py with the following options:

  • experiment_dir: Path to the output stored by LONER containing all logs relavent to the experiment. LONER reports this directory when it terminates.
  • traj_eval_dir: Where to store the processed files (for example, ./ablation_study/)
  • dataset: A name identifying the dataset. After this is set, output files will be placed in <traj_eval_dir>/<dataset>. This is useful for analyzing multiple runs, one per dataset.
  • groundtruth_traj_path: A path to the groundtruth trajectory in TUM format. This must start with timestamps at 0.
  • --single_config (flag, optional): Set this if there was only one configuration (i.e. --overrides was not set or had only one option)
  • --single_trial (flag, optional): Set this if only one trial happened per configuration (i.e. --num_repeats was not used or set to 1)

A common use-case is to test the algorithm on several datasets as follows:

python run_loner.py ../cfg/fusion_portable/<canteen/garden/mcr>.yaml --overrides ../cfg/ablation_study.yaml --num_repeats 5

Then you'd prepare results like this:

python3 prepare_results.py ../outputs/<canteen/garden/mcr>_<timestamp> ./ablation_study <canteen/garden/mcr> <path_to_groundtruth>

Running Analysis

Run cd ./ablation_study then source ~/LonerSLAM/analysis/compute_metrics/traj/analyze.sh, where ./ablation_study was generated above.

If you used a number of trials other than 5, you'll have to change the top line of analyze.sh accordingly.

Summarizing Results

Once the previous step worked, all that's left is to dump the information to a human-readable format. First, modify the following four variables in summarize_results.py:

  1. configs: This should match the list of configurations within <traj_eval_dir>/<dataset>, and will commonly be <config_0>,...,<config_N>.
  2. config_names: This is arbitrary, and only useful for human-readability. Setting algnames=algs is a reasonable choice if you don't want to set other names.
  3. datasets: This is the list of datasets. It should match the folders in <traj_eval_dir>
  4. metrics: Choose any combination of the metrics reported by evo_traj. Just ape is usually all you need here.

The results will be stored in csv and tex files in <traj_eval_dir>.

python compute_metrics/traj/summarize_results.py ./ablation_study

Map Metrics

The map metrics are trickier, and the workflow is even worse.

Computing L1 Depth

Run python3 analysis/compute_l1_depth.py <experiment_directory> to do the computation. The following options are available:

  • --single_threaded: By default, this will parallelize across all GPUs. If this flag is set, will only use one GPU.
  • --ckpt_id: If for some reason you don't want to use the final checkpoint, you can set this.
  • --num_frames: How many frames to compute the L1 depth for. Defaults to 25.
  • --use_est_poses: If set, uses LONER's estimates of pose rather than groundtruth poses.

The results will be stored in a yaml file: <experiment_directory>/metrics/l1.yaml.

Computing PointCloud Metrics

First, compute a mesh for all of the trials. I use this command, using canteen as an example (change to config for the dataset you tested on). This assumes your working directory is LonerSLAM/analysis/compute_metrics

for f in $(find <experiment_dir> -name "trial*"); do python3 ../meshing.py $f ../cfg/fusion_portable/canteen.yaml --skip_step 1 --resolution 0.05 --save --level 0.05; done

Refer to meshing.py for details on the options.

Next, convert the meshes to pointclouds. Resolution is generally .01 or .05.

for f in $(find <experiment_dir> -name "*.ply"); do python3 maps/mesh_to_pcd.py $f <resolution>; done

Finally, compute the metrics. You need to call evaluate_lidar_map.py on each of the files generated from mesh_to_pcd.py - it's pretty easy to write bash scripts for this. The arguments to evaluate_lidar_map.py are as follows:

  • experiment_directory: The experiment directory output from LONER. This needs to point to the individual trial, not the top-level directory.
  • gt_map: A pointcloud that is groundtruth
  • --gt_trajectory: Recommended: If set, uses this to compute an initial alignment between the groundtruth and estimated maps. TUM format.
  • --estimated_map: A path to the pcd generated by mesh_to_pcd.py
  • --est_traj: If set, uses this to help compute an initial alignment. Defaults to assuming this starts at the identity, which is generally good enough.
  • --initial_transform: If not using --gt_trajectory and --est_traj, uses this 4x4 homogenous transform as an initial transform to apply to the estimated map. This can be found via rough alignment in CloudCompare or similar.

Summarizing Results

From within analysis/compute_metrics/maps, first run prepare_results.py. The arguments are exactly the same as in the section on trajectory evaluation. Note that in this case, experiment_directory is the top-level output from LONER, not an individual trial. Next, skip straight to running summarize_results.py, exactly as in the trajectory evaluation section.