CommandsForPaperResults.md

Commands for obtaining the paper results

Download paper results for comparison:

You can download the paper results at: https://vision.in.tum.de/webshare/g/dm-vio/dm-vio_paper_results.zip

The plots and tables from the paper can be reproduced using the script paper_evaluations.py. For this you need to insert the path to the downloaded results at the beginning of the script.

Non-realtime results

These are not shown in the paper, but as they do not depend as much on the individual machine it makes sense to run them first.

python3 run_dmvio.py --output=null --dataset=euroc --dmvio_settings=euroc.yaml
python3 run_dmvio.py --output=null --dataset=tumvi --dmvio_settings=tumvi.yaml
python3 run_dmvio.py --output=null --dataset=4seasons --dmvio_settings=4seasons.yaml

Realtime paper results

python3 run_dmvio.py --output=null --dataset=euroc --dmvio_settings=euroc.yaml --realtime --dmvio_args="maxPreloadImages=16000"
python3 run_dmvio.py --output=null --dataset=tumvi --dmvio_settings=tumvi.yaml --realtime --dmvio_args="maxPreloadImages=16000"
python3 run_dmvio.py --output=null --dataset=4seasons --dmvio_settings=4seasons.yaml --realtime --dmvio_args="maxPreloadImages=16000"

Notes:

• adjust maxPreloadImages according to the RAM available on your machine, these values were meant to be used with 16GB. With enough RAM (>=32GB) you should not need this argument at all.
• for 4Seasons --dataset=4seasonsCR was used for the main paper results to make absolutely sure it uses exactly the same images as the other evaluated methods. But --datset=4seasons should be mostly the same and doesn't need the long dataset preprocessing (which can be performed by passing --crop_images to download_4seasons.py).

Method ablation (Fig. S1)

These should not be run in realtime mode as some of the ablations are not designed for it.

# Normal method
python3 run_dmvio.py --output=null --dataset=4seasons --dmvio_settings=4seasons.yaml
# 1. No Reinit and no Marginalization Replacement
python3 run_dmvio.py --output=null --dataset=4seasons --dmvio_settings=ablations/4seasonsNoReinitAndMargReplacement.yaml
# 2. No Initial Readvancing (+ all changes in 1.)
python3 run_dmvio.py --output=null --dataset=4seasons --dmvio_settings=ablations/4seasonsNoInitialReadvancing.yaml
# 3. No PGBA (+ all changes in 2.)
python3 run_dmvio.py --output=null --dataset=4seasons --dmvio_settings=ablations/4seasonsNoPGBA.yaml

Note that after the paper publication the parameter setting_minFramesBetweenKeyframes has been added, which slightly improves the non-realtime results and brings them closer to the realtime results. You can also reproduce the original paper ablation by adding --dmvio_args="setting_minFramesBetweenKeyframes=0" to all the lines above.

Weight ablation (Fig. S2)

# Disable dynamic weight
python3 run_dmvio.py --output=null --dataset=tumvi --dmvio_settings=tumvi.yaml --realtime --dmvio_args="maxPreloadImages=16000 dynamicWeightRMSEThresh=1e6"

Notes on reproducing the results:

Like DSO, DM-VIO is nondeterministic, which is why we run multiple times on each sequence. Therefore it is best to inspect the cumulative error plots (line_plot), rather than looking at individual executions.

The realtime results depend on the power and OS of the used system, which is why we recommend to first generate the non-realtime results. By default, in realtime mode the system will preload all images. Depending on the RAM available you should set the argument maxPreloadImages.