ruili3/pointdit_occ

PointDiT global-uniform-512d12

Image-to-point-cloud flow matching model (global_image_point_joint): frozen DINOv3 ViT-B/16 image encoder + 512-dim, 12-layer point decoder (153.6M params), 8192 points per image, trained on ScanNet++ img_full_gt (721 scenes / 71,712 images). Code: experiments/global_point/ on branch global-uniform-512d12 of the PointDiT repo.

Files

• global-uniform-512d12/checkpoint-last.pth — final checkpoint (epoch 15, step 16,815). Full training checkpoint: model (incl. frozen DINOv3), decoder_ema, optimizer, config, RNG states; can be passed to --resume. sha256 f71698c5e8a8e6aa2e6b7a087dec9180dcf068318a321786b0d77b3263e50b35.
• global-uniform-512d12/eval/eval-global512-{raw,ema}-steps{3,10,50}/ — metrics.json (protocol + image/scene/per-scene means) and visuals/: 135 sampled val images (first/middle/last image of each of the 45 scenes) with the input RGB (*.rgb.jpg) and GT + predicted point clouds (*.points.js, base64 float32 xyz). Open visuals/index.html locally (after downloading the folder) to view them. Each eval folder also has per_image_json.tar (4,479 per-image metric JSONs; metrics only, no point clouds) and eval.log.
• global-uniform-512d12/train/ — resolved_config.json and train.log (machine paths replaced by $RUN_ROOT/ and $CODE/), provenance.json, audit JSONs, tensorboard events.

Training

4×A800, global batch 64, lr 1e-4 (1 warmup epoch), AdamW wd 0.05, 15 epochs, uniform t sampler, noise scale 2.0, EMA 0.9999, bf16 autocast.

Evaluation (protocol official_val_raw100_global_v4)

ScanNet++ official val, 4,479 images / 45 scenes (no overlap with training scenes), bf16, seed 20261007, 8192 points, no test-time alignment. Image mean over all images (scene mean is nearly identical, see metrics.json). cd2, acc, comp: lower is better; F@τ: higher is better; radius_ratio: predicted/GT mean radius (1 = correct scale).

weights	steps	cd2	acc	comp	F@0.02	F@0.05	F@0.1	radius_ratio
raw	3	0.1844	0.1793	0.2929	0.030	0.138	0.263	0.430
raw	10	0.0701	0.1428	0.1283	0.045	0.256	0.506	0.802
raw	50	0.0620	0.1375	0.1122	0.048	0.267	0.551	0.961
ema	3	0.0902	0.1581	0.1595	0.037	0.216	0.433	0.694
ema	10	0.0827	0.1423	0.1564	0.039	0.230	0.470	0.924
ema	50	0.1050	0.1421	0.2014	0.035	0.183	0.393	0.996

Best: raw weights, 50 steps. EMA (decay 0.9999) has seen only ~17k steps and gets worse at 50 steps.

The RGB images and GT point clouds in visuals/ are derived from ScanNet++ and are subject to the ScanNet++ terms of use.