Readme

This repo demonstrates how to compute the covariances using MC dropout for the semantic keypoints in OrcVIO.

how to train

• use this repo
• pretrained weights are here

how to run

• modify the path

    root = '/home/erl/moshan/other_stuff/star_map_semantic_keypoints/'
    model_path = '/home/erl/moshan/orcvio_gamma/orcvio_gamma/pytorch_models/starmap/trained_models/with_dropout/model_cpu.pth'
    img_path = root + 'images/car2.png'
    det_name = root + 'det/car2.png'

• run the main

python src/main.py

• sample output

modifications to make it work with python 3

• modify pytorch

run

atom /lib/python3.6/site-packages/torch/serialization.py

and modify the code in _load

unpickler = pickle_module.Unpickler(f)
unpickler.persistent_load = persistent_load
result = unpickler.load()

change to

try:
    unpickler = pickle_module.Unpickler(f)
    unpickler.persistent_load = persistent_load
    result = unpickler.load()
except:
    unpickler = pickle_module.Unpickler(f, encoding='latin1')
    unpickler.persistent_load = persistent_load
    result = unpickler.load()

this is because the model is saved in python 2 but we use python 3

• modify mhParser

In src/mhParser.py, change the original part to this

for i in range(size // 2, det.shape[0] - size // 2):
  for j in range(size // 2, det.shape[1] - size // 2):
    pool[i, j] = (max(det[i - 1, j - 1], det[i - 1, j], det[i - 1, j + 1],
                     det[i, j - 1], det[i, j], det[i, j + 1],
                     det[i + 1, j - 1], det[i + 1, j], det[i + 1, j + 1]))