ego_gen_comind_single — CoMind SINGLE-EGO ablation dataset

Ablation of the CoMind 2-ego generation model where each view is generated independently as a single-ego stream (no cross-view information). Same 12,090 pairs / two ego views (leader L, helper H) as the main ego_gen_comind, but with three conditions changed:

1. No reference frames — reference appearance/pose/camera conditioning is removed entirely.
2. Own-context-only warping — *_warped is built from each view's OWN past pool only. Frames the main (combined-pool) model had to source from the partner pool (~11%) become holes (black frame + visibility all-False), i.e. the honest single-ego behaviour when the wearer's own past cannot cover the target.
3. Self-only pose — *_pose shows only the wearer's OWN hands (partner body removed). Identity palette kept (leader = EGO cyan/purple, helper = PARTNER orange/green) to match the main model's person_pose.
4. Per-view Plücker cameras — *_w2c is canonicalised to each view's own frame-0 (w2c[0] = identity), not the shared pair frame. Fed as camera_w2c, the model's shared canonicalisation inv(view0-frame0) becomes a no-op, so each view stays in its own coordinate = per-view Plücker (no cross-view shared world). role-mix safe.

Layout

clips/<rec8>/<clip_id>.single.npz     # self-contained training unit (below)
manifest/single_clips.txt             # the 12,090 clip_ids in this set
manifest/clip_pairs.csv               # pairing + metadata (superset; filter by single_clips.txt)
manifest/captions/<rec8>/<clip_id>.json  # per-file (SHARDED by recording: HF caps a dir at 10k files)
manifest/captions.json                # also provided: single bundle {clip_id: {"leader","helper"}}

<clip_id>.single.npz keys (all self-contained — no dependency on the main clip.npz)

key	shape	meaning
L_target, H_target	(77,) obj JPEG	GT video to generate (per view)
L_warped, H_warped	(77,) obj JPEG	own-only warped condition (hole = black frame)
L_vis_packed, H_vis_packed, vis_shape	packbits (T,504,504)	warp visibility (hole frame = all False)
L_srcf, H_srcf	(77,) int32	source past-frame index (hole = -1)
L_pose, H_pose	(77,) obj JPEG	self-only pose (own hands; L=EGO, H=PARTNER)
L_w2c, H_w2c	(77,4,4) f32	per-view canonical world→cam (frame0 = identity)
K_leader, K_helper	(3,3) f32	intrinsics (504 res)
warp_mode, pose_mode, cam_frame, n_hole_L, n_hole_H	—	provenance markers

Loader contract (needs a single_ego mode; NOT wired in the public loader yet)

For each clip read ONLY <clip_id>.single.npz:

• video ← {side}_target
• control_input_pose ← {side}_pose
• control_input_warped ← {side}_warped
• control_input_visibility ← unpack({side}_vis_packed, vis_shape)
• camera_w2c ← {side}_w2c (per-view canonical; feed directly, model canon becomes no-op)
• camera_K ← K_leader / K_helper, camera_src_res = 504
• caption ← manifest/captions/<rec8>/<clip_id>.json where rec8 = clip_id.split("_")[0] (per-file, sharded by recording — HF caps one dir at 10k files so they are NOT flat); or use the bundle manifest/captions.json[clip_id]
• do NOT load refs / reference_frames / reference_pose / reference_cam_w2c, and do NOT enable enable_reference_plucker. Everything else (net, person-pose embedder, Plücker embedder) is unchanged.