Reference

If this dataset is useful for your work, please cite our paper.


@article{dshppc2020,
title={DeepSemanticHPPC: Hypothesis-based Planning over Uncertain Semantic Point Clouds},
author={Han, Yutao and Lin, Hubert and Banfi, Jacopo and Campbell, Mark and Bala, Kavita},
journal={ICRA},
year={2020}
}

Dataset

Our dataset construction is here.

Download coco panoptic train/val 2017.
images_to_keep.txt lists training images of unstructured outdoor scenes (see paper)
If interested, mscoco-train2017-resnet50-places365.txt lists output of Places365 classifier
Convert segmentation labels using navigation_outdoor_categories-coco_panoptic_to_new_cats.txt
Human-readable category merging details given in navigation_outdoor_categories-merge_details.txt
Converted segmentation label id to label name given in navigation_outdoor_categories.txt

Network

Details to reproduce the network used in our paper are given here. For our network architecture, we use DeepLabv3+ with Xception65 backbone augmented with dropout in the middle and exit flow blocks for semantic segmentations. ASPP atrous rates are set to 6,12,18 for output stride 16 and 12,24,36 for output stride 8 as in \cite{deeplabv3plus2018}. Training uses output stride 16 while inference uses output stride 8. We initialize from network pretrained on ImageNet + MSCOCO + Pascal VOC, and train for 160000 iterations using SGD with momentum with batchsize 4 on our dataset. Initial learning rate 0.01, momentum 0.9, and polynomial learning rate decay with power 0.9 are used. Images are resized to 513x513.