predict_tile doesn't add image_path to boxes df

[youngdjn]

Describe the bug To save the predicted tree boxes to shapefile using boxes_to_shapefile, the boxes df produced by predict_tile needs to have a column image_path, but it doesn't. I'm using the recent update intended to fix this, but it's still not working. (The updated code looks good to my inexperienced python eye so I'm stumped.) I can manually add the image path with boxes["image_path"] = os.path.basename("/path/to/ortho.tif") and then boxes_to_shapefile works.

To Reproduce I insalled deepforest from github after the PR was merged: pip install git+https://github.com/weecology/DeepForest.git

To use the orthomosaic I have been using, download it from here: https://ucdavis.box.com/s/f6ddi0hzvtd5hq11bgp9l91fyjaydwlf

Then run my full script but change the ortho directory to where you downloaded it to:

from deepforest import main
from deepforest import get_data
from deepforest import utilities
import os
import rasterio as rio
import numpy as  np

r = rio.open("/home/exouser/Downloads/ept-ortho-crop.tif")
df = r.read()

rolled_df = np.rollaxis(df, 0,3)

m = main.deepforest()
m.use_release()

boxes = m.predict_tile(image=rolled_df, patch_size=1500)

# It only works if I also run this:
# boxes["image_path"] = os.path.basename("/home/exouser/Downloads/ept-ortho-crop.tif")

shp = utilities.boxes_to_shapefile(boxes, root_dir="/home/exouser/Downloads", projected=True)
shp.to_file("/home/exouser/Downloads/boxes.gpkg")

Environment (please complete the following information):

• OS: Ubuntu22
• Python version and environment : python 3.9

$ pip freeze | grep deepforest
deepforest @ git+https://github.com/weecology/DeepForest.git@d2de708bfcb2c170dd166def36988c7870c9a3c1

[Om-Doiphode]

I would like to work on this issue. Can you guide me on how to get started with it?

[bw4sz]

Yes, this additional line of code is needed, but i'm not convinced that's the problem. I think its that boxes_to_shapefile should take in the path to the RGB object and forget about the silly root dir (#512). I think this would be more explicit.

from deepforest import main
from deepforest import get_data
from deepforest import utilities
import os
import rasterio as rio
import numpy as  np

raster_path = get_data("OSBS_029.tif")
r = rio.open(raster_path)
df = r.read()

rolled_df = np.rollaxis(df, 0,3)

m = main.deepforest()
m.use_release()

boxes = m.predict_tile(image=rolled_df, patch_size=1500)

shp = utilities.boxes_to_shapefile(boxes, rgb=raster_path, projected=True)
shp.to_file("/home/exouser/Downloads/boxes.gpkg")

This way the boxes_to_shapefile function isn't making any assumptions about what columns are in 'boxes' for finding the rgb image. Let's close this issue and make a new one focused on that.