Skip to content

2017 07 10 Seventh meeting

Arthur Silber edited this page Jul 10, 2017 · 1 revision

Arthur:

  • Ported SSD to support python3 and mxnet 0.10
  • Created docker containers with python3, cuda and mxnet{master, 0.10, 0.9.5} built from source
  • Working on VisualBackpropagation (currently, shape of the labels do not match)

Fabian + Martin:

  • Working on mxnet amalgation (does not work yet)

Adrian:

  • Resnet ergebnisse: many prediction results with very high confidence (needs careful, high thresholds), but usually, the highest prediction result is the correct one
  • Merged "official" resnet architecture with our ssd, current training with resnet50 running
  • Discussion:
    • What result values does google get for different resnets (in their official paper)? - in general, resnets did not yield great results. Either use mobilenet for speed or inceptionnet for accuracy

Torben:

  • Added funkturm to "weaken" fernsehturm
  • But still, many false positives for fernsehturm
  • Added siegessäule: recognized very well
  • Added more images to none class: dublin spire, stuttgart fernsehturm, ..
  • Overall results: fernsehturm up to mAP 0.8 (from originally ~0.5), but false "fernsehturm positives" did not go down very well

Discussion:

  • Proposal Torben and Arthur: Focus on (1) Android App (2) Visual Backprop and maybe (3) binarized mxnet

Talk:

  • monday 15:15

  • 15min

  • similar to first talk, but show complete results

  • less focus on related work

  • idea

  • releated work

  • our progress

  • differences vgg resnet etc

  • dataset

  • Android app amalgation

  • (1) datensatz - updated pipeline(?) (2) network architecture (3) app

  • beispiel siegessäule zum datensatz hinzufügen?

ToDo for Paper:

  • Android App
  • Visual Backprop
  • Binarized(?) - optional