whuhxb / text2pos-cvpr2022

Code, dataset and models for our CVPR 2022 publication "Text2Pos"

Text2Pos: Text-to-point-cloud cross-modal localization

Code, models and KITTI360Pose dataset for our paper "Towards cross-modal pose localization from text-based position descriptions" at CVPR 2022. Main website: text2pos.github.io

Abstract

Natural language-based communication with mobile devices and home appliances is becoming increasingly popular and has the potential to become natural for communicating with mobile robots in the future. Towards this goal, we investigate cross-modal text-to-point-cloud localization that will allow us to specify, for example, a vehicle pick-up or goods delivery location. In particular, we propose Text2Pos, a cross-modal localization module that learns to align textual descriptions with localization cues in a coarse-to-fine manner. Given a point cloud of the environment, Text2Pos locates a position that is specified via a natural language-based description of the immediate surroundings. To train Text2Pos and study its performance, we construct KITTI360Pose, the first dataset for this task based on the recently introduced KITTI360 dataset. Our experiments show that we can localize $65%$ of textual queries within $15m$ distance to query locations for top-10 retrieved locations. This is a starting point that we hope will spark future developments towards language-based navigation.

Dependencies

Our core dependencies are

easydict
numpy
sklearn
matplotlib
cv2
open3d
plyfile
torch
torch_geometric

with pptk as an optional package to visualize the point clouds. Please pay close attention to the combined installation of CUDA, PyTorch and PyTorch Geometric as their versions are inter-connected and also depend on your available GPU. It might be required to install these packages by hand with specific versions instead of using the Pip-defaults.

The KITTI360Pose dataset

We create KITTI360Pose from the KITTI360 dataset. In particular, we sample 14,934 positions and generate up to three descriptions for each, totaling in 43,381 position-query pairs. We use five scenes (districts) for training (covering in total 11.59 km^2), one for model validation, and three for testing (covering in total 2.14 km^2). An average district covers an area of 1.78 km^2. Our baseline version of KITTI360Pose can be accessed HERE

Download the KITTI360Pose baseline

The baseline version can be found as a gzip file under our download link.

Create KITTI360Pose with custom parameters

After accessing the original KITTI360 and saving it under ./data/kitti360, our KITTI360Pose baseline for a specific scene can be created using

python -m datapreparation.kitti360pose.prepare --scene_name 2013_05_28_drive_0000_sync --cell_size 30 --cell_dist 10 --pose_dist 10 --pose_count 4 --shift_poses --grid_cells

These parameters can be varied to create other setups of the dataset, as used in our ablation studies.

Get started

Downloading pre-trained models

We make our pre-trained baseline models publicly available HERE. To run the evaluation, save them under

./checkpoints/coarse_contN_acc0.35_lr1_p256.pth
./checkpoints/fine_acc0.88_lr1_obj-6-16_p256.pth

Run evaluation

After setting up the dependencies and dataset and downloading our pre-trained models, the baseline evaluation can be run with:

python -m evaluation.pipeline --base_path ./data/k360_30-10_scG_pd10_pc4_spY_all_nm6/ \
--path_coarse ./checkpoints/coarse_contN_acc0.35_lr1_p256.pth \
--path_fine ./checkpoints/fine_acc0.88_lr1_obj-6-16_p256.pth 

Train baseline models

After setting up the dependencies and dataset, our baseline models can be trained using the following commands:

python -m training.coarse --batch_size 64 --learning_rate 0.001 --embed_dim 256 --shuffle --base_path ./data/k360_30-10_scG_pd10_pc4_spY_all_nm6/
python -m training.fine --batch_size 32 --learning_rate 0.0003 --embed_dim 128 --shuffle --base_path ./data/k360_30-10_scG_pd10_pc4_spY_all_nm6/