README.md

metadata

pretty_name: Ego-Exo4D develop (1-3 min takes, moving-object masks, metric depth)
task_categories:
  - depth-estimation
  - image-segmentation
tags:
  - egocentric
  - ego-exo4d
  - aria
  - metric-depth
  - moving-object-segmentation

Ego-Exo4D develop

The 1,659 Ego-Exo4D v2 takes of 60-180 s that have Aria RGB, both Aria SLAM videos and MPS calibration (44.1 h), with the raw files needed to compute depth, moving-object masks for all of them, and per-frame metric depth for 615 of them. Code that produced the masks and the depth, and runs on this folder as it is: https://github.com/XuweiyiChen/vlm_3d_tracking_annotator/tree/egoexo4d_depth/egoexo4d_depth

Layout

takes.json                                       Ego-Exo4D v2 take metadata (all takes of the release)
takes/<take>/frame_aligned_videos/
    <aria>_214-1.mp4                             Aria RGB, 30 fps, 1408x1408 fisheye (stored turned 90 degrees)
    <aria>_1201-1.mp4, <aria>_1201-2.mp4         Aria SLAM left / right
    cam01.mp4 ...                                exo GoPro videos (only 100 of the takes)
takes/<take>/trajectory/
    closed_loop_trajectory.csv                   MPS device poses
    online_calibration.jsonl                     MPS online calibration
    semidense_points.csv.gz, semidense_observations.csv.gz    MPS semi-dense point cloud and its observations
    gopro_calibs.csv                             exo GoPro calibration
captures/<capture>/timesync.csv                  capture timesync (frame-aligned video frame -> device timestamps)
gmos_masks/<take>/                               moving-object masks, all 1,659 takes
depth_mpsr/<take>.npz                            metric depth, 615 takes

The raw files keep the official Ego-Exo4D v2 layout and names (https://ego-exo4d-data.org).

Moving-object masks (gmos_masks/<take>/)

GMOS run on the whole take at 10 fps, one id per object for the whole take.

• instances.npz: instances uint8 (T, 512, 512), instance id per pixel, 0 = none; moving bool (T, max id + 1), GMOS's per-frame motion judgement of each instance (column 0 unused).
• cameras.npz: K (3x3) of the 512x512 view, frame_ids (frame index in the 30 fps frame-aligned videos; mask t is video frame 3t), image_valid (512x512, inside the lens circle), fps, source_fps.
• manifest.json: take, take_uid, camera, activity, task, frames.

The view: every 3rd frame of <aria>_214-1.mp4, turned upright and rectified to a 512x512 pinhole (focal 0.8 x the calibrated one, the take's middle online-calibration record).

Metric depth (depth_mpsr/<take>.npz)

Per-frame camera z-depth in metres for every pixel of the same 512x512 view (static scene and moving objects), at 10 fps over the longest run of frames that have MPS poses. Static scene: MoGe-3 corrected per frame to the MPS semi-dense points (plane-preserving 4-parameter fit by RANSAC), FoundationStereo on two Aria camera pairs, multi-frame vote and a 1 cm TSDF. Moving objects (all pixels of the GMOS instances): MoGe-3 times a local FoundationStereo ratio, refined by optimizing AllTracker tracks.

• depth: float32 (N, 512, 512), metres, 0 = no depth.
• frame_ids: int64 (N,), frame index in the 30 fps frame-aligned videos (mask index = frame_id // 3).
• K: 3x3 pinhole intrinsics; T_world_cam: float32 (N, 4, 4), camera to MPS world (OpenCV axes, metres); rgb_valid: (512, 512), pixels inside the RGB fisheye.

import numpy as np
d = np.load("depth_mpsr/cmu_bike01_4.npz")
i = 0
depth, K, T = d["depth"][i], d["K"], d["T_world_cam"][i]
v, u = np.mgrid[:512, :512]
points_cam = np.stack([(u - K[0, 2]) / K[0, 0], (v - K[1, 2]) / K[1, 1], np.ones((512, 512))], -1) * depth[..., None]
points_world = points_cam @ T[:3, :3].T + T[:3, 3]
masks = np.load("gmos_masks/cmu_bike01_4/instances.npz")["instances"][d["frame_ids"][i] // 3]

Recomputing

With the code above: python run.py --data <this folder> --out <output> --take <take>. To reuse the masks, copy gmos_masks/<take>/instances.npz and cameras.npz into <output>/<take>/masks/ and create <output>/<take>/masks.done.