jpena-173/sm-eyes-v2

Gaze-in-Wild (GIW)

Infrared eye-tracking video dataset captured in unconstrained real-world conditions at 640×480 @ 120 fps. This upload is a rearrangement of the original GIW video files and their corresponding GW (GazeinTheWild) per-frame annotations, reorganised into a unified directory tree for ease of use.

Parquet dataset

data/*.parquet contains every third frame (stride 3, starting at frame 0) from each eye video together with all GW per-frame annotations merged into a single table. The full 608 k-frame dataset yields 202,833 rows across 12 shards (one per video stream).

Loading

from datasets import load_dataset
ds = load_dataset("username/giw-gaze", split="train")

The image column is a HuggingFace Image feature (decoded automatically to a PIL image). Landmark columns are flat float32 lists; reshape as needed, e.g. np.array(row["lid_lm_2d"]).reshape(34, 2).

Schema

Column	Type	Description
subject	int8	Subject index (1–6)
session	int8	Session index within subject
eye	int8	0 = left eye, 1 = right eye
frame_idx	int32	0-based frame index (multiples of 3)
image	Image	640×480 IR eye frame (JPEG)
gaze_x/y/z	float32	Unit gaze direction vector in camera space
pupil_angle/cx/cy/w/h	float32	Pupil ellipse (angle °, centre px, axes px)
iris_angle/cx/cy/w/h	float32	Iris ellipse
pupil_in_iris_angle/cx/cy/w/h	float32	Pupil ellipse re-fit within the iris region
eye_ball_radius/x/y/z	float32	3-D eye-ball sphere fit
eye_movement	int8	1 Error · 2 Blink · 3 Fixation · 4 Saccade · 5 Smooth Pursuit
validity_pupil/lid/iris	bool	Annotation reliability flag
lid_lm_2d_inaccuracy	float32	Mean landmark inaccuracy (px)
lid_lm_2d	list[float32] × 68	34 eyelid landmarks in image space, flat [x₀,y₀,…] (17 upper + 17 lower)
iris_lm_2d_inaccuracy	float32
iris_lm_2d	list[float32] × 16	8 iris landmarks in image space
pupil_lm_2d_inaccuracy	float32
pupil_lm_2d	list[float32] × 16	8 pupil landmarks in image space
lid_lm_3d_inaccuracy	float32
lid_lm_3d	list[float32] × 102	34 eyelid landmarks in 3-D, flat [x₀,y₀,z₀,…]
iris_lm_3d_inaccuracy	float32
iris_lm_3d	list[float32] × 24	8 iris landmarks in 3-D
pupil_lm_3d_inaccuracy	float32
pupil_lm_3d	list[float32] × 24	8 pupil landmarks in 3-D

Missing annotations (validity = 0, or annotation absent for a frame) are stored as null.

Source layout

{subject}/{session}/
    eye0.mp4                   # left-eye IR video (640×480, 120 fps)
    eye1.mp4                   # right-eye IR video
    eye{N}.mp4gaze_vec.txt     # 3-D gaze direction unit vector (X Y Z)
    eye{N}.mp4pupil_eli.txt    # pupil ellipse (angle, cx, cy, width, height)
    eye{N}.mp4lid_lm_2D.txt   # 34 eyelid landmarks in image space (17 upper + 17 lower)
    eye{N}.mp4pupil_lm_2D.txt # 8 pupil landmarks in image space
    eye{N}.mp4iris_eli.txt     # iris ellipse parameters
    eye{N}.mp4iris_lm_2D.txt  # iris landmarks in image space
    eye{N}.mp4eye_movements.txt# per-frame eye movement classification (see below)
    eye{N}.mp4validity_pupil.txt
    eye{N}.mp4validity_lid.txt
    eye{N}.mp4validity_iris.txt
    eye{N}.mp4{iris,lid,pupil}_seg_2D.mp4      # 2-D segmentation mask videos
    eye{N}.mp4{iris,lid,pupil}_seg_3D.mp4      # 3-D segmentation mask videos
    eye{N}.mp4{iris,pupil}_seg_3Dplane.mp4
    eye{N}.mp4{iris,lid,pupil}_lm_3D.txt
    eye{N}.mp4eye_ball.txt

Annotation filenames mirror the GW convention (GW_{subject}_{session}) but are renamed to match the video they annotate (eye{N}.mp4{suffix}).

Coverage

Subject dir	GW subject	Sessions	Videos	Frames
1	GW_1	1	2	131,317
2	GW_12	1	2	109,234
3	GW_17	2	2	84,110
4	GW_18	1	2	79,641
5	GW_19	1	2	117,162
6	GW_20	1	2	87,022
Total		6	12	608,486

Annotation formats

All text annotations are semicolon-delimited with a one-line header and one row per frame.

gaze_vec.txt — FRAME;X;Y;Z — unit gaze direction vector in camera space.

pupil_eli.txt — FRAME;ANGLE;CENTER X;CENTER Y;WIDTH;HEIGHT — pupil ellipse fit in image pixels.

lid_lm_2D.txt — FRAME;AVG INACCURACY;LM X;LM Y;… — 34 eyelid landmark coordinates (17 upper then 17 lower) in image pixels.

eye_movements.txt — FRAME;Eye Movement Type — per-frame classification:

• 1 = Error
• 2 = Blink
• 3 = Fixation
• 4 = Saccade
• 5 = Smooth Pursuit

validity_{pupil,lid,iris}.txt — FRAME;VALIDITY — 1 if the annotation for that frame is reliable, 0 otherwise.

License

MIT — see LICENSE. Please cite the original GIW and GW papers if you use this dataset in your work.