What: 285 pick episodes (one complete pick attempt each, about 60 minutes in total) cut from 161 egocentric clips of real strawberry harvesting. Hand, head and chest mounted phone cameras, real field conditions, successes and failures both kept.
Annotation, per clip: 33 columns on the recording itself: camera and video properties, the picker's hands and pace, container, scene, lighting and weather, plus 6 edge-condition flags.
Annotation, per episode: outcome and recovery class, three event timestamps (grasp, detach, deposit), a measured occlusion profile, conditions of the target berry, an instruction string and the picker's narration.
In this repo: the tables, the 285 episode videos with audio, and the 161 uncut source recordings under strawberry-dataset-clips/.
Start here: data/episodes_wide.csv, one flat table, 285 rows by 58 columns.

The episode, one complete pick attempt from approach to resolution, is the training unit, not the clip.

Terms

Provided for evaluation and research. Please do not redistribute the data or pass access on; anyone who wants it can request access on this page. For anything beyond evaluation, contact doop.

Layout

strawberry-dataset-v1/
  README.md               this file
  data/episodes.csv       285 rows, one per episode (timestamps, labels, occlusion,
                             per-pick conditions, instruction, narration): 26 columns
  data/clips.csv          161 rows, one per source clip (capture context, environment,
                             6 edge flags): 33 columns
  data/episodes_wide.csv  the 285 episodes with their clip's columns joined in, 58 columns
  data/narration.csv      636 rows, the picker's spoken narration, one row per utterance
                             (episode_id, start/end in the file's timeline, verbatim text)
  episode-index.xlsx      285 rows, one per delivered video file (plain table)
  videos/                 285 mp4 files, named <episode_id>__<label>.mp4
  manifest.sha256         a checksum for every delivered file

Three tables, two levels. clips.csv describes each full recording as a whole: the camera, the light, the picker's habits over the clip. episodes.csv describes each individual pick cut from those recordings: what happened, when, and how visible the berry was. episodes_wide.csv carries both, one row per episode with its clip's columns joined in, so you only need one file.

Join key: uid (doop_001 to doop_179; gaps are clips curated out) links episodes to their source clip; episode_id (doop_XXX_eN) links a row in episodes.csv to its video file, one to one. Episode numbers come from the full annotation of each clip, in which non-action stretches (walking, searching between picks) were also numbered; those are not delivered, so eN sequences can have gaps. Merged arcs carry ids like doop_168_e1-e2.

Episodes and video files

Every episode file starts before the grasp, so the full approach is part of the episode and file durations routinely exceed grasp to deposit. Files carry lead-in and up to 1.5 s of tail where the source clip allows; in 7 episodes (doop_039_e5, doop_076_e3, doop_091_e3, doop_106_e2, doop_113_e2, doop_123_e2, doop_179_e1-e3) the deposit sits on the file's final frame. Adjacent episodes from the same clip may overlap slightly by design.
A recovery arc is ONE episode. Slip-regrasp and wrong-pick-correct arcs run from the first grasp to the final resolution and are never split.
Audio: natural field audio and the picker's spoken narration are kept.
Files are H.264 / AAC, faststart. Source capture was 4K or 1080p at 30 or 29.97 fps (per-clip details in clips.csv); each file keeps its source frame rate, 122 clips at exactly 30 fps and 39 at 29.97 (B_fps), so frame index is about round(t x fps).

Columns in clips.csv (one row per source recording)

uid: clip identifier, the join key to episodes.
A_gps_location: capture position as +lat+lon (ISO 6709 style, WGS84), the phone's own geotag. A_device: recording phone model.
B_viewpoint_mount: hand / head / chest, the perspective a viewer can verify from the footage, not the rig. B_orientation: how the footage renders (portrait / landscape). B_resolution: 3840x2160 or 1920x1080 as recorded. B_fps: measured source frame rate (30 / 29.97). B_codec: source codec (hevc / h264); delivered files are re-encoded H.264 regardless. B_clip_length_s: measured source duration. B_stability: stable / some-shake / very-shaky. B_quality_issue: dominant image defect (none, glare, motion-blur, autofocus-hunt, comma-joined).
C_hands: one-right / one-left / two. C_handedness_shown: right / left. C_pace: slow-careful / normal / fast-sloppy.
episodes_in_clip: number of delivered episodes cut from this clip (1 to 5). D_container_placement: punnet / basket / yes-unspecified / none. D_container_fill: empty / partial / full / na. D_narration_present: yes exactly when narration_quality is clean or partial.
E_mixed_ripeness_in_frame: ripe and unripe berries visible together (y / n).
F_background_clutter: clean / cluttered / people-in-frame.
H_light: midday-harsh / dappled / overcast-soft / golden / backlit / glare. H_time_of_day: morning / midday / afternoon / evening. H_weather: sunny / cloudy / overcast.
J_scene_berry_density: sparse / moderate / dense.
narration_quality: clean / partial / none. language: en where narration exists, blank otherwise. edge_flag_count: number of True edge_* flags on the row.
edge_*: 6 boolean flags, one line each under Edge cases.

Columns in episodes.csv (one row per pick)

Identity: episode_id, uid (the source clip), video_file, duration_s (file length in seconds).

Outcome and classes:

outcome: success (267), crushed (16), left-deliberately (2: the picker took hold of the berry, judged it not worth picking and left it on the plant).
decision: deliberate choice content of the episode: choose-among-cluster (30), reject-unripe (25), reject-damaged (6), reject-overripe (4), none. Independent of recovery_type.
recovery_type: slip-regrasp (42), placement-retry (36), wrong-pick-correct (31), crush-over-force (19), drop-miss (2), none (155).
- slip-regrasp: the fingers close on the berry and lose it while it is still on the plant, then take it again.
- placement-retry: the berry comes off cleanly but the first deposit misses or lands beside the container; the picker retrieves and places it.
- wrong-pick-correct: fingers close around the wrong berry, release it, and the correct pick follows. Touch-only is decision content, not a correction.
- crush-over-force: berry crushed by over-force. Outcome is crushed, except in 3 episodes where the picker crushes the berry, then picks another one cleanly in the same episode (outcome success).
- drop-miss: the grip fails in transit, the berry falls, is retrieved from the ground and deposited. Thin class (n=2).
technique_error: stem-loss (11): berry picked but the stem or calyx tore off.
grasp: pull-body (71), pinch-stem (25), twist (2), unknown (187, hand and stem not clearly visible).
container: punnet (159), basket (7), yes-unspecified (67, container visible but type unclear), not-visible (52, no container in the shot: the berry is put away out of frame or kept in the picking hand).
target_ripeness (ripe / under-ripe) and mixed_ripeness_in_frame (y / n / unknown).
target_touching_other_berry (single / touching) and unripe_berry_near_target (yes / no: a green or white berry within one berry width), both at the grasp. A value is present only where two annotators independently agreed (115 and 132 episodes); a blank cell means no decision, nothing was guessed.

Event timestamps, in seconds of the video file's own timeline:

field	event
`video_t_grasp`	first frame in which the fingers close on the berry that is later picked. A grip that slips and is retaken on the same berry counts from the first closure; a grip on a berry that is put back does not count
`video_t_detach`	the berry separates from the plant (in placement-retry arcs the original detach, never the re-pickup of a dropped berry)
`video_t_deposit`	the berry is released into the container or the picker's stash hand

Accuracy. Performance was evaluated against frame-by-frame human readings on a held-out sample of episodes: median error 0.50 s for the grasp, 0.45 s for the detach and 0.33 s for the deposit. Episodes flagged by the automated annotation layers were additionally verified by a person. The grasp is the least precise of the three, because the first closure is often hidden by the hand itself.

Blank cells are semantic, not missing data. An episode with no deposit (crushed, left deliberately, or the recording ends first) has no video_t_deposit. A berry crushed on the plant without detaching, or an attempt abandoned before the berry comes off, has no video_t_detach (doop_031_e4, doop_053_e2, doop_058_e2, doop_149_e1). An event that happens off film has no timestamp (doop_031_e6 begins with the berry already in the hand). In doop_010_e3, doop_039_e2 and doop_103_e2 the pick succeeds but the file ends before the deposit, so video_t_deposit is blank.

Occlusion, measured per pick. How visible the target berry was during the reach, taken as the last three seconds before the grasp plus grasp to detach, sampled at four frames per second. The picker's hand and forearm count as hiding the berry, and any visible part of the berry counts as visible.

occlusion_fraction: share of those frames in which the target berry cannot be seen at all. Median 0.25 across the dataset.
occlusion_fraction_low, occlusion_fraction_high: the annotator's stated range around the value. A wide range means the target was hard to follow, typically because the camera swung or several look-alike berries sat together.
target_mostly_hidden: yes when occlusion_fraction is 0.33 or more (118 episodes), no otherwise. A convenience flag; filter on the fraction for a different threshold.
occlusion_source: what hid the berry most: hand (the picker's own hand or forearm, 172), plant (leaves, stems, other berries, straw or wire, 44), offframe (the berry left the camera's field of view, 14), unidentifiable (the frames do not allow telling which berry is the target, 16), unclear (hidden, cause not attributable, 10), none (never hidden, 29).

Accuracy. Performance was evaluated against frame-by-frame human readings on a held-out sample of episodes: median absolute error on the fraction 0.087, mean 0.098. Episodes flagged as uncertain by the automated annotation layers were additionally verified by a person.

Text and quality:

instruction: one clean instruction string per episode, generated from the annotation (41 distinct strings).
narration_raw: the picker's spoken narration, verbatim (English), transcribed from the delivered files and spot-checked by a person. data/narration.csv has one row per utterance with start and end times, split at pauses of at least 0.4 s.
deliverable_quality: A (213) / B (72). A: every event the outcome requires is on film and timestamped, the action is visible, the camera is not very shaky and the clip has no autofocus hunt. B: any of those missing. Every delivered episode is A or B; non-action stretches were graded lower and are not delivered.
viewpoint: hand / head / chest, as B_viewpoint_mount above.

Constants (true for every clip, stated once instead of shipping constant columns)

Lens 1x on all clips. Straw mulch, ground-row field culture, medium row spacing, low plants, dry ground in all delivered clips. Variety: June-bearing type, fruiting stage. Bare hands in every clip.

Capture context

Multiple field sites, distinguishable via per-clip GPS. Real commercial harvest conditions. Handedness: right 124 / left 37 clips. Lighting skews midday-harsh and sunny with dappled and overcast minorities (H_* columns).

Concurrent recordings

In 44 cases two or three phones were recording at the same time in the same field, each on a different picker doing their own picks. These are not the same pick seen from two cameras. What the clips in a group share is the moment, so light, weather and field state are identical across them, and the other picker is sometimes audible.

Edge cases (clip-level flags, 6 boolean columns)

Each flag marks a condition that makes perception or control harder and says it occurs somewhere in the recording. Per-pick conditions (occlusion, an unripe neighbour, a dropped berry) are per-pick columns in episodes.csv or follow from the recovery class.

edge_another_person_in_frame: a second person is visible in the shot.
edge_backlit_silhouette: hand or target reduced to a silhouette by backlight.
edge_dappled_moving_light: moving leaf-shadow speckle plays over the scene.
edge_ground_obstacle_pipe: an irrigation pipe or similar obstacle lies in the working area.
edge_rain_on_lens: raindrops on the lens distort the image.
edge_wet_glare: wet surfaces throw specular glare.

edge_dappled_moving_light is common (81 clips); the other five are rare.

Known limitations

Thin classes: drop-miss n=2, stem-loss n=11, crush-then-success n=3.
grasp type is unknown for 187 of 285 episodes.
All footage is rights-cleared and purpose-shot by consenting contributors. A second person appears in frame in 3 clips (flagged edge_another_person_in_frame). Redaction concerns: contact doop.

Loading

import pandas as pd
eps   = pd.read_csv("data/episodes.csv")        # 285 rows, one per video file
narr  = pd.read_csv("data/narration.csv")       # timestamped narration segments
clips = pd.read_csv("data/clips.csv")           # capture context, join on uid
wide  = pd.read_csv("data/episodes_wide.csv")   # the two joined, ready for EDA

Source clips layer (full recordings)

The complete source recordings ship in this repository under strawberry-dataset-clips/videos/: the 161 clips (9.2 GB, 71.8 min, 4K/1080p HEVC originals), one file per clips.csv row, named <uid>__<original camera name>.MOV. The episodes in videos/ are windows cut from these recordings; the full clips add the between-pick context (walking, searching, repositioning) as continuous, uncut footage. Narration outside delivered episode windows is not transcribed in narration.csv. manifest.sha256 inside the folder lists a checksum per file.

Downloads last month: -

Total file size:

13.3 GB