AnyCam_DL3DV_480p

Object-centric camera-control training clips built from DL3DV videos. Each sample is a pair of 81-frame windows from the same scene (source and target), each with a control video of one anchor object:

• target condition: the anchor object reconstructed as a Gaussian splat from one frame, aligned to the scene and rendered alone along the target window's ground-truth camera trajectory, on white.
• source condition: the real anchor object cut out of the source video with its SAM 2 mask, on white.

The Wan 2.1 VAE latents of the same samples are in AnyCam_DL3DV_480p_wan21_latent.

Version 2 (2026-10-05). This replaces the first release (3,462 samples, 2026-09-28). Changes: the splats are rendered with the DL3DV camera intrinsics (the first release used SAM 3D's focal estimate, which is 17-20 % too long, so renders were too small); the source condition is the SAM 2 cut-out of the real object instead of the splat render; windows and selection were redone.

Contents

• 5221 samples (tier good: 4222, tier poor: 999), 4 videos each.
• All videos: 832x480, 81 consecutive frames, 30 fps, H.264 (crf 16).

metadata.jsonl                 one row per sample
<scene_id>/target.mp4          target video
<scene_id>/target_cond.mp4     splat render of the anchor object along the target window
<scene_id>/source.mp4          source video
<scene_id>/source_cond.mp4     SAM 2 cut-out of the anchor object in the source window
<scene_id>/meta.json           the sample's metadata row

scene_id is dl3dv_<sha256>, the DL3DV scene hash. Videos are a centre crop (x 10-842) of the 852x480 re-encode of the DL3DV video.

How the samples were made

1. A VLM (Qwen3-VL-32B) picks one rigid anchor object per scene.
2. SAM 3 grounds it, SAM 2 tracks it through the video.
3. SAM 3D Objects reconstructs the object as a Gaussian splat from one frame.
4. The splat is aligned to the DL3DV camera trajectory (1-DOF scale search, then a 7-DOF SIM(3) refinement with an IoU loss) and rendered alone with the DL3DV camera intrinsics.
5. Two non-overlapping 81-frame windows are cut per scene. Windows avoid detected camera-pose jumps and need the object visible in the render in every frame.
6. Inside each window SAM 2 re-tracks the object, guided by the render's position. The source window must have a mask in every frame and a mean mask-vs-render IoU of at least 0.5; otherwise source and target are swapped or another window is used.
7. Selection: every sample was scored by a model trained on manual keep/drop judgments, using only measurements inside the two windows. Samples scoring >= 0.90 were kept automatically (selection == "score", 3023; blind spot checks kept 97-100 %); the rest were reviewed by hand and only the ones judged good were kept (selection == "manual", 2198).

Special cases, all marked in the metadata:

• source.mode == "static" (926 samples): the scene has only one usable window. The source is a single frame repeated 81 times, a frame that shows the whole object.
• padded == true (target 276, source 83): the window has 70-80 usable frames; the last frame is repeated up to 81.
• Source-condition frames where SAM 2 has no mask of the object (e.g. it is hidden) are white.

Metadata fields

field	meaning
scene_id	DL3DV scene
variant, intrinsics	alignment variant (iou) and render intrinsics (camera = DL3DV K)
tier	good / poor scene-level quality tier
prompt	scene caption (describes the whole scene, independent of the anchor object)
object	anchor object phrase chosen by the VLM; in a few samples SAM 3 grounded a different object than the phrase names
target, source	window: start frame in the scene video, len, padded, complete, visible, path and rot_deg (camera motion), area (object area fraction); source.mode = window / static
source_cond	sam2_cutout
source_fix, source_clean	present when the source window was swapped or re-cut (step 6)
n_camera_jumps	detected camera-pose jumps in the whole scene (windows avoid them)
longest_run, usable_total	frames with a visible render available in the scene
selection, score	score (kept automatically) or manual; the selection score
files	relative paths of the four videos

Source data

Derived from DL3DV-10K. Use of this dataset is subject to the DL3DV-10K terms of use.