I am fine-tuning detector models that almost work in the target domain, but not quite. In particular, I'm fine-tuning MegaDetector (MD) for cases where it struggles, but still has some signal. In these cases, if you use MD to generate boxes, you will get many target objects and save yourself gobs of time in building a training set, but you will miss some target objects entirely, and you will have to use a sufficiently low confidence threshold that you get a bunch of junk. Before I do any labelme work, I do an aggressive repeat detection elimination pass, but still, there's some junk, and some misses, but mostly good boxes, and I want to clean all that up to make a new training set.

I compared a few OSS labeling tools and found that none quite supported this scenario, since the common path is still the one where you're making boxes from scratch. But labelme was the easiest to populate with bounding boxes from ML results, and by far the easiest for me to modify (all in Python, no fancy-schmancy Web infrastructure).

I wanted to add a few UI features specific to the scenario where boxes are prepopulated from detector-generated boxes, to generate data I could use to train new detectors. This scenario has a few unique UI requirements:

• Deleting boxes efficiently (i.e., without using the mouse) is more important than in the typical de-novo-boxes scenario
• Being able to definitively mark images as empty is important
• Fine adjustment of close-but-not-quite-there boxes (with the keyboard) is important
• Being able to page quickly through mostly-correct boxes and see huge bright red boxes that require almost no cognitive processing time is important

• New keyboard shortcuts (defaulting to the integers 0-4) to toggle labels starting with numeric integers, e.g. "0" toggles the first label starting with the string "0".

• Alt-PageDown advances to the next non-empty image (and ditto for Alt-PageUp)

• Alt-I toggles the "ignore" flag (if present)

• Ctrl-PageDown and Shift-PageDown advance by 100 and 20 images, respectively (configurable in the .yaml file), ditto for PageUp. Related: the current index in the file list is now displayed in the title bar.

• Clicking an empty area in the canvas, then shift-clicking another point, selects all rectangles that intersect with the rectangle defined by the two points you just clicked.

• Deletion no longer raises a confirmation prompt (every click counts!)

• Shift/ctrl + up/down/left right move the upper-left and lower-right borders of a selected rectangle.

• PageUp and PageDown select the previous/next image, in addition to the default A/D shortcuts. This is useful when you're moving really fast, and you want to switch hands for the next/prev action. (This turned out to be unnecessary, I just didn't know you could already map multiple shortcuts to the same action. Now I know. But pgup/pgdn are still hard-coded in my fork.)

• The save action is enabled by default (so we can save images with no boxes)

• A new action (default K) to keep only the selected polygons

• A new action (default M) to merge all shapes in an image into one rectangle that's the union of the current shapes (useful when an object has been split into multiple overlapping objects)

• A new action (default B) to keep only the largest ("Biggest") rectangle

• A new action (default S) to keep only the smallest rectangle

• A new action (default shift-B) to select the largest ("Biggest") rectangle

• A new action (default shift-S) to select the smallest rectangle

• A new action (default L) to load labels from image_file.alt.json, instead of image_file.json, typically used to load a version of the pregenerated labels that uses a lower confidence threshold

• Variations on that action (default Alt-0 ... Alt-9) to load from image_filt.alt-N.json instead

• A new field ("saved_by_labelme") is written to the output on every save, and saving happens every time you change images; together, these allow us to confirm that an image has been reviewed, even if no changes were made.

• Shortcut to copy the current file to the clipboard (useful for re-starting where you left off when you've prepopulated the list, so the checkboxes aren't useuful)

• Allow saving annotations to the output_dir when output_dir is specified (I don't really remember why I did this)

• Keyboard shortcut to select all polygons

• Command-line argument to resume from the last image you saved

• Customizable line width, including a command line option --linewidth, because when paging through images at five images per second, it's really helpful to have huge, bright boxes.

• ...but thin boxes and lower opacity are helpful when doing fine adjustment, so the selected box has a thin line and reduced opacity.

• Changes to the defaults that make it faster for cases where you're mostly just confirming boxes: a brighter color for the "animal" class, auto-save by default, don't save image binary data to .json by default

• Alt-right and alt-left to select next/prev boxes (super-useful for the case where multiple overlapping boxes are predicted for a single object, which is a time-consuming situation to resolve with the mouse)

cd ~/git
git clone https://github.com/agentmorris/labelme
cd labelme
mamba create -n labelme-git python=3.11 pip -y && mamba activate labelme-git && pip install -e .

python labelme folder_name --labels animal --last_updated_file ~/labelme-last-updated.txt

python labelme folder_name --labels animal --last_updated_file ~/labelme-last-updated.txt --resume_from_last_update

labelme --reset-config

...because I got QT errors.

sudo apt-get upgrade -y
sudo apt install -y libgl1-mesa-dev
# I don’t think this was necessary
export QT_QPA_PLATFORM="xcb"
sudo apt install libxcb-xinerama0 libqt5x11extras5

• A,D or PgUp,PgDn (previous/next)
• L (load the alternative annotation file for this image) (.alt.json)
• Shift-L (load the backup alternative annotation file for this image) (.alt-0.json)
• Ctrl-L (load the backup-backup alternative annotation file for this image) (.alt-1.json)
• M (merge all rectangles into one rectangle that's the union of everything)
• K (keep only selected polygons)
• B (keep only the largest ("B" for "biggest") rectangle)
• S (keep only the smallest rectangle)
• Ctrl-R (create rectangle mode)
• Ctrl-J (edit polygons mode)
• Ctrl-A (select all polygons)
• Ctrl-E (edit the label of the selected polygon)
• Alt-right/alt-left (select next/previous polygons)
• Shift/control + left/right/up/down (fine adjustment of 0th/1st vertices)
• Delete (delete current polygon)
• Ctrl-C,ctrl-v (copy/paste selected polygons)

• My fine adjustment logic breaks down a little when boxes are near the edge of the canvas (nothing bad happens, fine adjustment just stops working), fix this