yslan/ECCV26_PhysAI_Challenge_NVS_Syn4D_subset

ECCV'26 PhysAI Challenge — Novel-View Synthesis (Syn4D subset)

A small, self-contained novel-view-synthesis (NVS) evaluation package for the ECCV'26 PhysAI workshop challenge (video-diffusion NVS under camera control). It ships the held-out Syn4D inputs (source rgb + depth + camera + segmentation mask) and the matching RecamMaster raw-resolution predictions, so you can reproduce the reference metrics end-to-end with no dataset access and no inference.

👉 The evaluation CODE lives here, not on this dataset page

Scorer / benchmark: https://github.com/NIRVANALAN/eccv26-physai-workshop-vdm_nvs_bench (vdm-nvs-bench) — install it, then score any submission (DAVIS or Syn4D track). This HF dataset is only the data package (inputs + a reference model's outputs).

---

What's inside

One file: nvs_syn4d_eval_set_recammaster.tar.zst (~311 MB). It expands to:

nvs_syn4d_eval_set/
  official/<pair_id>/          # 7 eval pairs — flying_group/seq_000001, src view 0 -> tgt views 1..7
      source.mp4 target.mp4         # RGB 81x480x832  (target = novel-view GT)
      cameras.npz                   # source/target/rel_target c2w (81,4,4) + K + frame_ids
      {source,target}_depth.npz     # z-depth (81,480,832) f16
      {source,target}_mask.npz      # dynamic-fg mask (¬env)
      meta.json + previews
  recammaster_out/<pair_id>/   # RecamMaster (original step20000) RAW predictions, 832x480x81
      pred_*.mp4 gt_*.mp4 source_*.mp4 compare_*.mp4
  tools/ run_recammaster_1gpu.sh docs/  score_out_reference/   # scorer adapter, runner, docs, expected metrics
  README.md  REPRODUCE.md

flying_group is one of two never-trained held-out Syn4D scenes. A→B semantics: source view 0 is the input; the model renders each target view's camera trajectory; target.mp4 is the real rendered target = paired GT.

Quick start

# 1) download this dataset and extract:
hf download yslan/ECCV26_PhysAI_Challenge_NVS_Syn4D_subset --repo-type=dataset --local-dir nvs_syn4d_subset
tar --use-compress-program=unzstd -xf nvs_syn4d_subset/nvs_syn4d_eval_set_recammaster.tar.zst
cd nvs_syn4d_eval_set

# 2) set up the scorer ONCE (see the GitHub repo's "Setup"): a python env, the OFFICIAL
#    vggt-omega, and the bench itself:
#      conda create -n vdm-nvs-bench python=3.10 -y && conda activate vdm-nvs-bench
#      git clone https://github.com/facebookresearch/vggt-omega && pip install -e vggt-omega
#      git clone https://github.com/NIRVANALAN/eccv26-physai-workshop-vdm_nvs_bench && \
#        pip install -e eccv26-physai-workshop-vdm_nvs_bench && \
#        python eccv26-physai-workshop-vdm_nvs_bench/scripts/download_weights.py

# 3) score (the bench is on PATH after pip install; else set VDM_NVS_BENCH=/path/to/it):
python tools/score_recammaster.py --out score_out
python tools/final_row.py

Full walkthrough (prereqs, re-inference, scoring your own model): REPRODUCE.md inside the archive.

Reference metrics (what you'll reproduce)

RecamMaster (original ReCamMaster step20000, raw 832×480×81), 7 pairs, canonical ViT-H-14 CLIP, native 81 frames:

n	ATE ↓	trans ↓	rot° ↓	CLIP-V ↑	CLIP-F ↑	PSNR ↑	SSIM ↑	LPIPS ↓
7	0.0478	0.0224	0.434	0.862	0.969	12.61	0.227	0.573

Camera = VGGT-Omega vs the requested trajectory; appearance = paired vs GT. FVD is omitted — 7 clips is far too small for a stable Fréchet estimate (it is still computed into score_out/video_metrics.json, just not headlined).

Reproduce the predictions from the official RecamMaster repo

The shipped recammaster_out/ predictions were generated by the original ReCamMaster step20000 checkpoint at native 832×480×81. To regenerate them yourself:

Prereqs: the recammaster-official repo (provides the diffsynth WanVideoReCamMasterPipeline), the original ReCamMaster step20000.ckpt + the Wan2.1-T2V-1.3B base, the Syn4D dataset, and its conda env (sync4d).

Option A — the official repo's own eval script (authentic RecamMaster baseline):

cd /path/to/recammaster-official
GPUS=0,1 SCENE=flying_group SEQ_ROOT_LIST=seq_000001 \
  SOURCE_VIEW=0 TARGET_VIEWS=1,2,3,4,5,6,7 \
  CKPT_PATH=/path/to/ReCamMaster/checkpoints/step20000.ckpt \
  DATASET_ROOT=/path/to/Syn4D \
  bash bash_scripts/eval-metrics/eval-syn4d-flyinggroup-recammaster-step20000-2gpu.sh

This samples all 7 targets at native 832×480×81, 81 frames (inference_recammaster_syn4d.py), splitting the targets across 2 GPUs. Outputs land under eval/syn4d-nvs-syn4dEval/multimodal_flying_group_seq_000001/recammaster_official_step20000_.../flying_group/seq_000001/ as pred_native_rgb_src0_to_tgt{1..7}_chunk0-80.mp4 (+ gt_native_rgb_*, source_native_view0_*).

Option B — the bundled 1-GPU runner (identical inference code, drops straight into the scorer layout — this is what produced the shipped recammaster_out/):

cd nvs_syn4d_eval_set
export RECAM_REPO=/path/to/recammaster-official      # provides diffsynth
export CKPT_PATH=/path/to/ReCamMaster/checkpoints/step20000.ckpt
export DATASET_ROOT=/path/to/Syn4D
GPU=0 N=7 FORCE_RERUN=1 bash run_recammaster_1gpu.sh   # writes recammaster_out/<pair>/…/pred_*.mp4

The RecamMaster inference entry script is vendored in the archive under tools/recammaster/ (only the heavy diffsynth pipeline comes from RECAM_REPO). Inference is deterministic (fixed seed=0) → the same ckpt reproduces the same frames. Then score as in Quick start.

License / attribution

Derived from the Syn4D dataset; released for the ECCV'26 PhysAI challenge under CC-BY-NC-4.0. Predictions are from the original ReCamMaster model. Please cite the challenge and the respective source works.