<redacted>
Trains are composed of an engine/locomotive and train "cars" joined by a disconnectable joint called coupler or a coupling. The purpose of this computer vision model is to detect these couplers from the image as viewed by a camera mounted on a pole next to a railway track. The usecase is not stated but the output is supposed to be x coordinate. Evaluting metric is stated as "accuracy".
- very small dataset (100)
- Images from very few locations, ligting conditions, etc.
- more types of coupler (coupler between bogies, on-bogie-coupler - unsure of terminology)
- multiple overlapping trains on multi-tracks
- the axis of the camera doesn't need to be perpendicular to the tracks
- camera can be in a place where there is a bend in the tracks.
- it might be crucial that the model is resilient against different types of weather and ligting conditions (eg. lighting angle change, night).
- looks like north-american railway system
- output format doesn't allow multiple couplings in image
- not common dataset format (eg. COCO's labels.json)
- missing image file
- crooked 4-point polygons instead of standard bounding box format
- large/inaccurate annotations encompasing more than coupling (sorrounding bogies)
- use a metric (maybe mAP)
- 1 channel model for B/W
- augumentations (albumentations)
- dropout
Poetry: poetry install, poetry shell
pip: pip install -r requirements, pip install -e .
- extract dataset to
car_coupling_train python ./coupler_detector/train.py(~5 CPU-core-minutes)
test random file:
./coupler_detector/find_couplings $(ls -1 car_coupling_train/*.jp*|sort -R|head -1) --preview
To get validation set image paths:
python ./coupler_detector/dataset.py
Edited example image by: Florent Dellesale
