yslan/holo4d_dataset_caption

Holo4D dataset captions: inventory, sampling, and use

Audit updated: 2026-09-24. This document describes the locally available captions at /vgg/yushi/dataset/caption_output/gemini_holo4d/ and the published subset in the Hugging Face dataset yslan/holo4d_dataset_caption. The caption repository contains annotations only; RGB, depth, calibration, poses, and 3D tracks remain in the source datasets and are loaded by the local unified dataloader.

Quick policy

• A local/window caption describes one temporally bounded clip. Join it to a training crop by annotation.sample_id and its start_frame/end_frame.
• A global caption describes one sequence using up to 64 frames distributed uniformly over its full duration. It is useful as sequence context, but is less time-specific than an aligned local caption.
• TartanAir and MVS-Synth already have 64-raw-frame local captions (nominal stride 32); they are not global captions and do not need a new global caption for the current training plan.
• Kubric49 and Kubric81 are short clips. Their local captions (49 frames for Kubric49; 64-frame windows for Kubric81) are the intended annotations; no additional global captions are required.
• Syn4D v1 and v2 have both local and global products. DL3DV currently uses global captions. OpenVidHD and OmniWorld-Game use the datasets' own annotations; do not replace those with Gemini captions.
• Never train on an error record. Captioning code writes "A scene is shown." as a fallback prompt when the provider/validator failed. That phrase is not a successful caption. Every successful Gemini record checked in this audit has zero occurrences of that placeholder. Filter raw JSONL on error is absent and a populated annotation, as shown below.

Current inventory and readiness

Counts below are record counts from the current local JSONL/official metadata. “Failed rows” are resumable-run diagnostics, not captions. Some runs append a later success for the same sample; unresolved counts compare by the precise window annotation.sample_id, not just by video ID.

Dataset	Caption type / current output	Successful records	Failed rows	Unresolved caption items	Training status
Syn4D v1	Local raw96, stride 96; syn4d_v1/captions_gemini-3.8-flash_holo4d_semantic_v3_n096_success.jsonl	15,638	Historical error rows have successful counterparts	0 local windows	Local windows ready. Separate global uniform64: 4,515 successes, 3 sequences still unresolved.
Syn4D v2	Local raw96: syn4d_v2/captions_gemini-3_8-flash_high_holo4d_semantic_v3_1_raw96_success.jsonl	10,127 / 10,128 planned	14 raw error attempts	1 policy-blocked window	Usable for all successful windows; one input was blocked by provider policy. Global uniform64: 4,799 / 4,800; same one sequence blocked.
DL3DV	Global uniform64: dl3dv/global_captions_gemini-3.8-flash_holo4d_semantic_v3_uniform64_success.jsonl	6,377	Historical failures excluded from success export	0 among remaining usable sequences	Global captions cover every remaining sequence. A truncated 12-frame corrupt source sequence was removed and is excluded.
MVS-Synth	Local raw64, nominal stride 32; mvs_synth/captions_gemini-3.8-flash_holo4d_semantic_v3_n064.jsonl	360	52	0 unresolved windows	120 sequences × 3 planned windows; all have a successful caption after retry. Local only is sufficient for this plan.
TartanAir	Local raw64, nominal stride 32; tartanair/captions_gemini-3.8-flash_holo4d_semantic_v3_n064.jsonl	9,384	18	0 unresolved windows	All planned windows have a successful caption after retry. Local only is sufficient.
ReCamMaster	Local raw64 at 15 FPS; recammaster/captions_gemini-3.8-flash_holo4d_semantic_v3_n064.jsonl	115,597	219	7 windows	Most local windows are ready; seven remain without captions (IDs below). No separate global JSONL is currently the canonical product.
Kubric49	Local raw49; kubric49/captions_gemini-3.8-flash_holo4d_semantic_v3_n049.jsonl	11,269	3	3 windows	Short clips need local only. Three end-aligned windows remain without captions; other windows are present.
Kubric81	Local raw64 windows; kubric81/captions_gemini-3.8-flash_holo4d_semantic_v3_n064.jsonl	7,118	93	0 unresolved windows	Short clips need local only. Every failed sample ID has a successful record.
OpenVidHD	Official local annotation CSV: /vgg/yushi/dataset/OpenVidHD_60k/annotations/local_openvidhd_63337.csv	63,328 CSV rows	—	9 of 63,337 local MP4s have no matching CSV row	Use the official captions/metadata; no Gemini pass needed.
OmniWorld-Game	Official frame-range JSON: /vgg/yushi/dataset/OmniWorld_Game/extracted/annotations/OmniWorld-Game/*/text/*.json	Text chunks found for 438 of 479 extracted sequence/UID folders	—	41 folders have no text JSON in this inventory	Use official per-chunk annotations and their frame ranges, not one synthetic trajectory-wide caption.

Remaining caption gaps

The following are annotation failures, not confirmed source-data corruption. Do not delete their videos just to make counts match. Retry them when the caption API is available, or explicitly exclude those samples from training.

• Syn4D v1 global (3): lab_bald/seq_000010_1, planet_bald/seq_000015_1, winter_group_static/seq_000013_3.
• Syn4D v2: one policy-blocked local window and its corresponding global sequence. The provider prohibited the input; do not repeatedly retry it.
• ReCamMaster local (7): f18_aperture10/scene778/cam05_f000017_f000080_n064, f24_aperture5/scene1309/cam01_f000017_f000080_n064, f24_aperture5/scene1309/cam02_f000017_f000080_n064, f24_aperture5/scene1309/cam03_f000017_f000080_n064, f24_aperture5/scene1309/cam07_f000017_f000080_n064, f24_aperture5/scene1882/cam10_f000000_f000063_n064, f24_aperture5/scene303/cam04_f000017_f000080_n064.
• Kubric49 local (3): train/00815/camera_2_f000001_f000049_n049, train/00823/camera_2_f000001_f000049_n049, train/01105/camera_3_f000001_f000049_n049.
• OpenVidHD: nine local files lack a corresponding row in the official annotation CSV. Treat those as missing annotations rather than inventing a placeholder.
• OmniWorld-Game: 41 extracted folders lack a text JSON in the current annotation tree; inspect the official UID/trajectory manifest before deciding whether they are truly missing or intentionally uncaptioned.

How the local captions are sampled

Dataset	Clip length and stride	FPS / meaning
Syn4D v1	96 consecutive raw frames, sample_stride=1; block stride 96, end-aligned final block if needed	30 FPS. Local caption describes that raw-frame block. A separate global record samples up to 64 frames uniformly over the full sequence.
Syn4D v2	96 consecutive raw frames, sample_stride=1; end-aligned last window if needed	30 FPS. Independent global caption samples up to 64 frames over the complete sequence (for a 192-frame sequence, the global record spans the full 192-frame extent).
MVS-Synth	64 consecutive raw frames; nominal stride 32; end-aligned final window	30 FPS. Example starts are 0, 32, then an end-aligned start such as 36 in a 100-frame sequence.
TartanAir	64 consecutive raw frames; stride 32; end-aligned final window	30 FPS. Example starts 0, 32, 64…; the last window may shift backward to end exactly at the sequence end.
ReCamMaster	64 consecutive raw frames; nominal stride 32; end-aligned final window	15 FPS. For an 81-frame clip, the starts are 0 and 17, so the final block reaches the last source frame.
Kubric49	49 consecutive raw frames; end-aligned window where needed	Short source clips; use this local caption rather than adding a global caption.
Kubric81	64 consecutive raw frames per local window; end-aligned final window	Short source clips; local-only policy. The exact selected block is in each record.
DL3DV	No canonical local-window release currently; one global record per remaining sequence	Up to 64 uniformly distributed observations across the whole sequence, not 64 adjacent raw frames.
OpenVidHD	Official dataset-provided caption	Preserve the upstream video/caption association in its CSV.
OmniWorld-Game	Official caption chunks with chunk-specific frame ranges and fields	Use the official chunk timestamps/frame ranges and map them onto decoded frames.

For a local caption, the training crop should stay inside its associated caption window. For global captions, a short training crop may use the sequence-level prompt, but that prompt can mention events outside the crop. Avoid silently substituting a local prompt for another window from the same video.

Random sample quality audit

One record was randomly selected and read for each caption product/dataset (Syn4D v1/v2 were checked separately for local and global; official captions were checked in their native format). The sampled Gemini captions were specific, scene-grounded descriptions, not generic filler:

Product	Example observed in the random spot-check	Result
Syn4D v1 local96	Indoor lab/room with a rug, people dancing and watching	Pass
Syn4D v1 global64	Desert settlement with people/figures and a fox	Pass
Syn4D v2 local96	Urban street/manhole setting with a person and nearby objects	Pass
Syn4D v2 global64	Courtyard with entities and actions; 64 samples spread over the full sequence	Pass
DL3DV global64	Park entrance/architecture	Pass
MVS-Synth local64	Urban street corner, pedestrians, parked car and fire hydrant	Pass
TartanAir local64	Rainy futuristic industrial/factory setting with a smoking vehicle	Pass
ReCamMaster local64	Indoor atrium with a person gesturing/raising an arm	Pass
Kubric49 local49	Sunlit tabletop with boxes/objects in motion	Pass
Kubric81 local64	Park lawn with objects tumbling/sliding	Pass
OpenVidHD official	Hamster eating; upstream caption is descriptive	Pass
OmniWorld-Game official chunk	Armored warrior fighting, then walking in a snowy forest; camera/action fields are present	Pass

This is a spot-check, not a guarantee that every caption is factually perfect. Automated validation found zero successful Gemini records containing a scene is shown across the checked exports. The literal phrase does occur in failed raw rows as a fallback; those rows have an error field and empty semantic annotations, and must be filtered out.

Record format and safe loading

Gemini outputs are JSONL: each non-empty line is one JSON object. The common fields are:

• vdm_prompt: compiled text suitable for VDM conditioning.
• annotation: structured holo4d_semantic_v3 content (scene, subjects, objects, relations, timeline, visible text, summary, uncertainty).
• metadata.video_id: portable source sequence key; use this to join to a dataset, not machine-specific absolute paths.
• annotation.sample_id: unique local window/global record identifier.
• annotation.window.start_frame, end_frame, num_frames: inclusive source-frame interval and number of observations.
• metadata.sampled_frames: exact raw frame indices supplied to the model.
• metadata.observation_scope: source_only for local clips or full_sequence for global descriptions.
• error: present on failed attempts. Such records are diagnostic only.

Safe local loading/filtering:

import json

def read_successful_captions(path):
    with open(path, encoding="utf-8") as f:
        for line in f:
            if not line.strip():
                continue
            record = json.loads(line)
            annotation = record.get("annotation")
            if record.get("error") is not None or not isinstance(annotation, dict):
                continue
            prompt = (record.get("vdm_prompt") or "").strip()
            if not prompt or prompt.casefold() == "a scene is shown.":
                continue
            yield record

for record in read_successful_captions("captions.jsonl"):
    caption_id = record["annotation"]["sample_id"]
    text_condition = record["vdm_prompt"]

Join and temporal alignment:

sample_id = record["annotation"]["sample_id"]
video_id = record["metadata"]["video_id"]
start = record["annotation"]["window"]["start_frame"]
end = record["annotation"]["window"]["end_frame"]
# Decode/select the training crop from this video_id within [start, end].

The semantic caption does not replace camera calibration, camera motion, depth, optical flow, 3D tracks, or geometry. Those modalities must come from the source dataset through the unified dataloader.

Hub access

The caption repository is yslan/holo4d_dataset_caption (dataset repo type). Example for a published success-only file:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="yslan/holo4d_dataset_caption",
    repo_type="dataset",
    filename=(
        "syn4d_v1/"
        "captions_gemini-3.8-flash_holo4d_semantic_v3_n096_success.jsonl"
    ),
)
with open(path, encoding="utf-8") as f:
    records = [json.loads(line) for line in f if line.strip()]

At this audit, the published clean exports include Syn4D v1 local96/global64, Syn4D v2 local raw96/global64, and DL3DV global64. The MVS-Synth, TartanAir, ReCamMaster, and Kubric Gemini files are present in the local VGG output tree; check the Hub file list before assuming those local files have been uploaded.

Official OpenVidHD and OmniWorld-Game annotations

OpenVidHD captions are in its local annotation CSV and are not in the Gemini semantic JSONL schema. The audited local CSV has 63,328 rows for 63,337 videos.

OmniWorld-Game JSON files have a top-level captions object. A sample chunk contains these keys:

Short_Caption, PC_Caption, Background_Caption,
Camera_Caption, Video_Caption, Key_Tags

The chunk filename/official metadata identifies its frame range. Keep those fields and ranges intact; the different caption types describe the action, point-character/actor, background, camera, and whole chunk. These official annotations are chunk-level, not one forced global caption per trajectory.

Corrupt-source handling

Only verified malformed/truncated source data should be removed. The known DL3DV sequence with only 12 frames was confirmed unusable and removed from the local source tree; its caption is consequently excluded. Provider quota, service errors, generic-language validator errors, or policy blocks do not prove that the source video is corrupt. Those cases are listed as caption gaps above and should be retried or explicitly excluded from training, not deleted.