README.md

metadata

language:
  - en
license: mit
pretty_name: False Absence Detection
tags:
  - clarusc64
  - world-models
  - spatial-grounding
  - absence
  - persistence
  - video
task_categories:
  - video-classification
  - image-classification
size_categories:
  - n<1K
source_datasets:
  - original

ClarusC64/false_absence_detection_v01

Dataset summary

This dataset tests whether models treat invisible entities as gone or still present.
In each sequence, a target leaves the camera view.
Some exits are real.
Some are false absences with clear evidence that the target remains in the container.

Goal

• check if models infer continued presence from indirect cues
• avoid treating every disappearance as an exit
• keep spatial grounding under occlusion and clutter

Key signals

• absence_tag: present, still_present, left_scene
• evidence_type: shadow, reflection, audio cue, motion trail, interaction trace, door state, physical constraint, none
• false_absence_target: true when the model should infer “still present”
• trap_flag: true when the sample is designed to lure models into assuming exit

Columns

• sample_id – unique id per frame sample
• split – train, valid, eval
• modality – video
• scene_type – indoor_room, factory_line, corridor, sports_pitch, warehouse
• sequence_id – id for a temporal sequence
• frame_index – index within the sequence
• time_gap – not used here (fixed gaps inside raw video)
• container_id – id of the main container
• container_bounds – "x_min y_min x_max y_max"
• boundary_type – hard, soft, porous
• zone_id – region id inside the container
• zone_type – sofa, door, conveyor, chute, passage, turn_left, exit_door, left_flank, stands, robot_lane, intersection, dock_exit
• target_entity_id – tracked entity such as cat_01, crate_05, cart_02, ball_07, robot_11
• target_visibility – visible, partial, not_visible
• absence_tag – present, still_present, left_scene
• evidence_type – none, shadow, motion_trail, interaction_trace, audio_cue, door_state, reflection, crowd_reaction, physical_constraint
• false_absence_target – true if the model should infer continued presence from evidence
• trap_flag – true if the scene is designed as a false-absence trap
• label_type – baseline, false_absence_case, true_exit
• drift_risk – low, medium, high
• comment – short human note

Example loading code

from datasets import load_dataset

ds = load_dataset("ClarusC64/false_absence_detection_v01")

row = ds["train"][1]
print(row["target_entity_id"], row["target_visibility"],
      row["absence_tag"], row["evidence_type"], row["false_absence_target"])