Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Apolloscape dataset for localization task.

Exploring localization task on Apolloscape dataset.

Read my blog post about PoseNet implementation details

ECCV2018 Self-localization on-the-fly challenge task details.

NOTE: This repository is a work in progress.


Dataset reader based on Pytorch 0.4.1 Dataset. To install all dependencies:

pip install -r requirements.txt


Download Apolloscape data from page and unpack it to a folder. Examples below assume that data folder symbolically linked to apolloscape-loc/data/apolloscape.

mkdir ./data
ln -s <DATA FOLDER>/apolloscape ./data

Sample data file for zpark road provided in localization challenge section supported automatically (it has different folder names, files order and pose data files format)

Python Notebook example

See roads and record graphs in Apolloscape_View_Records Notebook

PoseNet training, error calculation and result visualization in Apolloscape_PoseNet

PoseNet on Train

Show/Save path and sample images by record id

python --data ./data/apolloscape --road road03_seg --record Record018

Generate video of the path by record id

python --data ./data/apolloscape --road road03_seg --record Record018 --video

Record video

Train PoseNet convnet on ZPark road

python --data ./data/apolloscape --road zpark-sample --checkpoint-save 50 --fig-save 1 --epochs 2000 --lr 1e-5 --experiment zpark_posenet_L1 --feature-net resnet34 --feature-net-pretrained --learn-beta

Training process:

Training and validation results:


  • VidLoc implementation
  • [Optional] Prepare data for eval script
  • SfM / 3D Reconstruction pipeline
  • WGAN for generating new samples
  • Qt/OpenGL visualizations