Each image is a single frame of one fish on a conveyor, with a YOLO bounding box. A subset have YOLO segmentation and
keypoints. Frames are grouped by species (ITIS code) and then by capture event; several consecutive frames are usually
available for each fish.
conveyor/
classes.json YOLO class id -> ITIS species
keypoints.json keypoint order, names and groups
length_weight.json NOAA length-weight parameters by species and month
LICENSE GPLv3 license text
NOTICE copyright notice
itis164744/ species folder (ITIS TSN)
0f3c…e2_5/ capture event
camera.json camera calibration
itis164744_38cm_0f3c…_20240808T015043.600000Z.jpg image
itis164744_38cm_0f3c…_20240808T015043.600000Z.txt box
itis164744_38cm_0f3c…_20240808T015043.600000Z.seg.txt outline
itis164744_38cm_0f3c…_20240808T015043.600000Z.key.txt keypoints
File and folder names are identifiers used to trace each frame back to its source event; treat them as opaque.
All label files use YOLO formats with coordinates normalized to 0–1 of the image width and height, one line per fish:
.txt box: class x_center y_center width height
.seg.txt outline: class x1 y1 x2 y2 … xn yn (polygon from a SAM 2.1 model prompted with the box; it may extend beyond the box)
- In the image viewer, "Outline image (.png)" downloads the fish cut out along its outline on a transparent background,
with the page's current "Enhance details" setting applied (0% gives the original pixels). "Image (.jpg)" is always the
unmodified original.
.key.txt keypoints: class x_center y_center width height followed by x y v for each of the 16 keypoints in
keypoints.json order (v = 2 labelled, 0 not labelled)
camera.json (one per event): camera model, camera_matrix and
distortion_coefficients (OpenCV conventions), plane_homography
from undistorted normalized coordinates to the belt plane in mm, camera_height_mm,
scale_px_per_mm on the belt, and a fish_height model for the outline's height above the belt
(a shape-based default plus a species adjustment, also listed in classes.json).
In the image viewer, the fish's center length and area are measured from its outline with these settings.
Species whose length_measure in classes.json is "fork" are measured
to the tail fork (where the tail lobes meet), all others to the tail tip.
- Outlines and keypoints created in the image viewer ("Create SI Outline", "Create SI Keypoints") are made by SI models in
your browser and are not part of the dataset. They can be downloaded as "SI" files named
<frame>.si.seg.txt, <frame>.si.key.txt and <frame>.si.outline.png.
Created keypoints that belong on the fish's edge (snout tip, fin and tail fin points) are placed on the outline.
- Species (SI), shown with the SI length, is experimental: a small image classifier run in your browser on the fish cut
out by its outline. It knows the species with at least 150 fish in the dataset and shows "(insufficient training
data)" for the others, "(unclear)" when no species is above 50% confidence, and otherwise the most likely species
(and a second one if it also reaches 30%). It was trained on about half of the fish of those species; these are marked * ("was included in
training data"), and the model will look more accurate on them than on new fish.
- Estimated weights use the NOAA NEFSC length-weight fits (Wigley et al. 2003) in
length_weight.json, by species and month of capture.
To train with Ultralytics YOLO, which reads <image>.txt from a labels/ folder, copy the label type you
want there under the image's name, e.g.
for f in */*/*.seg.txt; do cp "$f" "labels/$(basename "${f%.seg.txt}").txt"; done.
For pose, add kpt_shape: [16, 3] and the flip_idx from keypoints.json
to your dataset yaml.
Class ids map to species in classes.json and in the species table.
Species descriptions follow the NOAA NEFSC
ITIS species codes
("NK" = not known to species level).