Inspiration

Most models provide no confidence measure for its predictions due to argmax(). Thats a problem if you want to compare the predictions for objects from different sensors (multimodalilty)

What it does

  • we colored the images with a different intensity proportional to the output of the neuron (plotted logits)
  • with a pre-build sotA model we predict the class of objects in images
  • with this classification we trained a separate convNN to predict errors in the classification.

How we built it

python (tensorflow) with google deeplab

Challenges we ran into

  • core of challenge was already popular research area, so we found it hard to identify possible contributions for this scope
  • modifying the current state-of-the-art model-graph to include dropout for inference (in order to achieve dropout sampling for confidence estimation) proved a technological challenge

Accomplishments that we're proud of

hacked dropout into model

What we learned

we learned to develop not perfect NN-models in short time.

What's next for detecting errors in pixelwise segmentation

Built With

Share this project:

Updates