2016 05 22 First meeting: research results

What has been done

Found Datasets (created wiki page here):

• Oxford Building Dataset: 5k images, semi-sorted, no BBs
• Oxford Paris Dataset: 500-1k images, sorted by landmark, no BBs
• Paris500k Dataset: 501356 images in total, cralled from flickr, 10 different building labels
• LabelMe: has per-pixel labels (-> BB extraction via script possible)
• MIT Scene Parsing Benchmark: semi-useful per-pixel labels (Buildings are often too big / cropped on images)

Discussion:

• Restrict images to dailight only? (many photos found on google images, e.g. Brandenburger Tor, are taken at night)
• Image synthesis - generate test images manually --> use for pre-training to aid model training --> Arthur: 3D renderings / Martin: photogammetry
• Fabian: found a thesis: memory efficient database for mobile image search --> research how its done, anything we can use? - Here is the link to the thesis: Memory-efficient image databases for mobile visual search - A link to his page: David M. Chen

Network Training approach:

1. General training e.g. imgnet
2. Train especially on "is there a building"
3. specific training e.g. for berlin (hand annotated images?)

Next steps

1. Reserach mxnet SSD in detail
2. How to build image dataset:
   • --> 6a) test segmentation - background versus building / search class "building" - test with images
   • how to generate BBs?
   • image synthesis
   • goal: pretraining
   • MaskRCNN?

Who:

• Adrian: Labels to BB (LabelMe dataset)
• Arthur: test SSD with pictures