To start using the annotator you just need to go to the dlc2action_annotation directory and
run these commands in your terminal.
conda activate dlc2action_gui
python annotator.pyThe app will prompt you to open the video you are going to annotate. It can be multiple videos and in that case you will be asked whether you want to open them sequentially or in multiview mode.
Alternatively, you can set the video and/or the output paths in the terminal. In that case the file dialogs will not open.
python annotator.py --video path/to/video1.mp4 --video path/to/video2.mkv --multiviewYou can modify the GUI configuration through the settings window. It will open by default the first time you run the interface and you can always go
back by adding an -s / --open_settings option when you run the app or pressing 'Settings' in the menu bar when it's already open.
Backups for your annotated data are automatically created every 30 minutes in a folder
located next to your first video: path/to/video1_backups. You can change the directory
where backups are saved for your project, and the interval at which they are saved.
Running
python annotator.py --backup-dir /path/to/backups --backup-interval 120will save backups every 2 hours (120 minutes) in the /path/to/backups folder.
The results of your work will be saved at path/to/video1_annotation.pickle. The _annotation.pickle
suffix is the default, you can change it in the settings window (at Files / Annotation suffix). If you
open a video that already has a corresponding annotation file in the same folder, that file will be loaded
automatically. A human-readable version of the annotations will also be saved in CSV
format (with the same suffix, and .csv extension).
After you start the app it might take up to a couple of minutes for the program to load if the video file is large. When it does you will be prompted to enter the labels you want to use. The right column is the shortcut that you can use to annotate that label while watching the video. It is automatically defined as the first letter of the label name unless that is already reserved, but you can always change it. You can come back to this dialogue later, even after you start the annotation, so it's okay if you don't enter everything (or don't enter anything at all) now.
If you provide a list of labels in the settings, the same dialog will open with pre-loaded categories for fine-tuning. When you're done with entering the labels, press OK. That should take you to the main window.
Let's go over the main elements here.
-
Video canvas
Drag and zoom are supported. If multiple animals are displayed and their keypoints have been uploaded, the animal that is being annotated is the one with the colored keypoint markers. -
Action bar
This bar zooms in on the frames closest to the one currently displayed. The thick gray line shows where you are at the moment. As soon as you start the annotation the labels will be displayed here. If an annotated clip is too short, the label will not be displayed on the bar. However, you can always see it in the status bar when you hover over an action. You can mark actions as ambiguous (eigher manually in the Ambiguous mode or during shortcut annotation, more on both options in the corresponding sections). In that case the actions will be transparent on the action bar (like jumping in this example). Also, clicking anywhere on the bar will take you to the corresponding frame. -
Global bar
Here you can see where you are in respect to the entire video and how much of it is currently in the working memory of your machine (blue color). The counter on the right displays the index of the current frame and the total number of frames in the video. This bar is also clickable, but keep in mind that after clicking outside the blue zone you may need to wait for the video to load. -
Status bar
The status bar displays contextual prompts. -
Speed sliders
You can use the video speed slider to adjust the speed of video playback. When keypoints are provided, displaying them might limit the speed. In order to avoid that you can adjust the speed of detection update. Setting it to zero will stop keypoint display altogether. -
Individuals menu
Pick the animal you want to annotate. You can also use shortcuts (numbers from 0 to 9). -
Labels menu
The labels and the shortcuts are displayed here. If you want to add a new action manually, you need to select the label for it in this menu. Use the 'Labels' menu to modify it. If you are using nested annotation, you can choose a category by double-clicking it and go back to the categories list by pressingEscor pushing the 'Go back to categories' button. -
Mode toolbar
Here you can see which of the manual modes is active right now. More on them in the manual annotation section. You can switch between the modes by pushing the corresponding button here or in the 'Manual' menu or by pressing a shortcut (see a cheatsheet in the 'Manual' menu). Note that the status bar always displays the relevant hints for any mode. The leftmost button on the toolbar starts and pauses the video. You can drag and drop the whole panel to organize the workspace the way you like. -
Playlist navigation
If you open multiple videos in sequential mode or use label search you will also see these buttons that allow you to move between videos or clips. Your annotation will be saved automatically when you move to a different video. -
Menu bar
You can find some additional modes and settings in the menu bar. Read on to find out more on some of them. Note that you can also access the settings window from here.
You can start the video by pressing spacebar or the 'Play/Stop' button in the toolbar. Alternatively, you can browse frame-by-frame by pressing the arrow buttons on your keyboard. If you press one of the shortcuts before browsing the video in one of these ways everything you see is automatically annotated with the corresponding label until you press the shortcut again. Multiple shortcuts can be active at the same time.
In case you are working with a long structured list of labels, it might be more convenient to use nested annotation. To do that just check that box in the 'Load from list' dialog you will see when the program loads.
In this mode you can start by annotating a category and specify the label later (either with a shortcut -- the
right menu will open automatically -- or manually in the Assign mode). You can open one of the categories by double-clicking it and go back to the category menu by pressing Esc.
You can switch between normal and nested annotation whenever you want: just open the 'Load from labels' menu and check/uncheck the box!
You can also just leave the category as the label, without specifying anything. Press Esc to go back to the categories menu and use the same shortcut to stop the annotation.
In addition to pressing the specific label shortcut (like 'L' for leaping), you can also stop annotating a label using the Stop mode. Just press '-' to enter it. One of the labels you are currently annotating will be highlighted. If you press Enter it will be stopped, '-' will move the selection to the next label and Escape will take you out of this mode. This might be useful if you have multiple shortcut menus, like in nested annotation.
You can also mark the labels you are annotating as ambiguous using the same mechanics, but with the '=' key. Pressing '=' will get you to the Ambiguous mode, Escape will take you back, '=' navigates through labels and Enter changes their ambiguity status. If it was certain it will become ambiguous and vice versa.
Note that the status bar always displays prompts in case you forget those shortcuts!
If you want to add an action manually, just select it in the categories menu, check that the mode is New (double-click an action / select the plus icon / press Ctrl + N) and
click and drag on the action bar.
You can always modify all actions by clicking and dragging the edges in the Move mode.
Clicking an action in the Remove mode (select the trash bin icon / press Cmd + R) deletes it.
Clicking an action in the Cut mode (select the scissors icon / press Cmd + C) splits it in two.
Clicking an action in the Ambiguous mode (select the transparency icon / press Cmd + B) changes its ambiguity.
Clicking an action in the Assign mode (select the label icon / press Cmd + A) changes its label to the one you choose in the label menu.
There are a couple of display options you can configure.
Press 'Bodypart colors' to color the keypoints according to their bodypart name (the outline will still be colored according to the individual).
You can also check which bodypart each keypoint represents by pressing it and checking the status bar.
If you want to get an overview of your annotation, you can use the 'Start label search' (or 'Start unlabeled search') option in the 'Active learning' menu. It will go over all the clips in this video file that you have annotated with a specific label. If you have loaded a keypoint file, the individual that was annotated with this label will be highlighted and all others will be grey. Press 'Next' and 'Previous' buttons on the control panel to navigate through the clips and select 'Start/Stop active learning' in the 'Active learning' menu to go back to the normal annotation mode.
Note : You might need to save the annotation file before starting the search !
The program should run smoothly, but please don't forget to save your results regularly by selecting the 'Save' action in the 'File' menu or pressing Cmd + S, just in case.
The app accepts videos in all commonly used formats.
The deeplabcut output files can be either tracklets (in .pickle format) or tracks (in .h5 format). You can also load keypoint sequences in the format of the MABE21 challenge (in that case set data format to calms21 in the general settings).
You can also use the interface to open 3d keypoint files as a separate view. To do that, set the 3d suffix in the file settings. If the video you are opening is at /folder/VIDEO.mp4, and the 3d suffix is set to _3d.npy, the app will expect the 3d files to be at /folder/VIDEO_3d.npy. The files should be numpy arrays saved in .npy format with shape (#frames, #keypoints, 3). The names of the corresponding bodyparts can be specified in the
display settings.
If you are loading a 3D pose file, you can also display reprojections on other views if a path to a calibration folder is provided in the file settings. The calibration folder is assumed to contain multiple .npy files containing information for different views:
calibration_folder
├── camera_aa_calibration.npy
└── camera_ab_caliration.npy
In this case the settings from the camera_aa_calibration.npy file will be applied to videos that have filenames starting with aa- and the settings from the camera_ab_calibration.npy file to videos that have filenames starting with ab-. Each of the .npy files should contain a dictionary with the following keys: "r" with the rotation matrix, "t" with the translation vector, "Intrinsic" with the intrinsic parameter matrix, "dist_coeff" with the distortion coefficients. The format of all parameters should be as specified in cv2.projectPoints'.
You can also open .msgpack segmentation masks. Each frame should be a list of dictionarirs that have "category_id" and "segmentation" keys. See utils.Segmentation class for more details.
Your annotation will be saved as pickled files. They can be opened with this python code.
import pickle
with open("/path/to/annotation.pickle", "rb") as f:
data = pickle.load(f)
metadata = data[0]
label_list = data[1]
individual_list = data[2]
annotation = data[3]To look up the annotation for individual "ind0", for instance, you can then run the following.
ind_annotation = annotation[individual_list.index("ind0")]
for label_index, label in enumerate(label_list):
for start, end, ambiguity in ind_annotation[label_index]:
print(f'from {start} to {end} frame: {label} label with ambiguity {ambiguity}')The ambiguity of an interval will be 1 if you marked it as ambiguous (transparent) and 0 if you didn't.
The program works best with shorter, smaller files. In case you want to cut your larger video and skeleton files into several pieces, you can use the
split.py script. Just open the classic_annotation folder in your terminal and run the following:
python split.py --file path/to/video --split-size S --downsample N --fps F --skeleton-file /path/to/DLCHere S encodes the size of each piece in seconds, downsample lets you only keep every Nth frame (useful in case of large fps) and F is the input video framerate. The skeleton-file parameter is optional.
You can find detailed explanation on how to use the clustering interface on this page.
Back to main readme