Scrape Instagram

[mattmotoki]

Scrape Instagram for images of plants on the following lists:

• Native Plants
• Invasive Plants
• Common Plants
  • hibiscus
  • plumeria
  • pineapple
  • ginger
  • anthurium
  • heliconia
  • bougainvillea

Upload all data to our shared images directory.

[xyl012]

scraped instagram for ~400 pictures in each hashtag. Includes non-plant pictures. Will go through images depending on how many needed for each class. Running Inception and mobilenet classifiers on noisy data. Out of box and with noisy data, inception v3 65-70% validation accuracy https://github.com/tensorflow/hub/blob/master/examples/image_retraining/retrain.py

https://drive.google.com/drive/folders/1TauCtJxGnEOYLX0_iKVXltuS2hNVNf3d?usp=sharing

[xyl012]

Deleted misclassified images, ~2600 images total in common_v2

[mattmotoki]

Nice, the images in common_v2 look a lot cleaner! Can you share the code you are using too (I'm assuming it's a bit more than just a straight application of the link you posted)? Can we apply whatever you are using to the native and invasive lists too?

[xyl012]

I went through all the pictures manually for V2 to make ground truth certain. The pictures are all from mobile/regular life so the settings are variable ranging from selfies to coffee shop commercials. I used this github to scrape posts with a free AWS (fastest internet) original instagram scraper. Only edited to get --maximum 500 posts and command line loop to get 7 hashtags. I'm not sure about the native/invasives, but I'm in the process of testing. The tensorflow retraining uses folders as classes so I don't need to create a csv, but it should be straightforward with ls >. Simple use of the example is here retraining as submodule

[mattmotoki]

Wow very nice! Sounds good. Thanks Chris!