Add final r-filter retention checkpoints (25/50/75%) and model card
Browse files- README.md +44 -0
- retention_r25_re10k.ckpt +3 -0
- retention_r50_re10k.ckpt +3 -0
- retention_r75_re10k.ckpt +3 -0
README.md
CHANGED
|
@@ -1,3 +1,47 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- 3d-gaussian-splatting
|
| 5 |
+
- novel-view-synthesis
|
| 6 |
+
- pose-free
|
| 7 |
+
datasets:
|
| 8 |
+
- re10k
|
| 9 |
---
|
| 10 |
+
|
| 11 |
+
# Compact NoPoSplat: stratified retention with a reliability filter
|
| 12 |
+
|
| 13 |
+
Checkpoints for a compact feed-forward 3DGS model built on
|
| 14 |
+
[NoPoSplat](https://github.com/cvg/NoPoSplat). The model renders with 25, 50 or
|
| 15 |
+
75% of the 131K Gaussians NoPoSplat predicts for two 256 × 256 views.
|
| 16 |
+
|
| 17 |
+
Code, method and reproduction steps: https://github.com/junyoung6215/capstone_project
|
| 18 |
+
(`docs/METHOD.md`, `docs/REPRODUCE.md`).
|
| 19 |
+
|
| 20 |
+
## Files
|
| 21 |
+
|
| 22 |
+
| File | Gaussians kept | #Gaussians | PSNR | SSIM | LPIPS |
|
| 23 |
+
| --- | --- | --- | --- | --- | --- |
|
| 24 |
+
| `retention_r25_re10k.ckpt` | 25% | 33K | 24.62 | 0.817 | 0.184 |
|
| 25 |
+
| `retention_r50_re10k.ckpt` | 50% | 66K | 25.01 | 0.829 | 0.169 |
|
| 26 |
+
| `retention_r75_re10k.ckpt` | 75% | 98K | 25.26 | 0.837 | 0.161 |
|
| 27 |
+
|
| 28 |
+
RE10K test set (5,590 evaluable scenes of the official index), pose alignment on.
|
| 29 |
+
NoPoSplat reports 25.03 / 0.838 / 0.160 with all 131K Gaussians.
|
| 30 |
+
|
| 31 |
+
Each model starts from the OLP-H + reliability head base checkpoint
|
| 32 |
+
(`kyoosuk/noposplat-olp-rel-re10k`, `epoch_1-step_75000.ckpt`) and is fine-tuned
|
| 33 |
+
for 10K steps with pruning at its own ratio: 2 × 2 stratified cells, r threshold
|
| 34 |
+
0.35, batch 1 × 4 accumulation, lr 2e-5.
|
| 35 |
+
|
| 36 |
+
The files hold model weights only (optimizer state removed); they load for
|
| 37 |
+
evaluation, not for resuming training.
|
| 38 |
+
|
| 39 |
+
## Evaluate
|
| 40 |
+
|
| 41 |
+
```bash
|
| 42 |
+
python scripts/run_retention.py --config my_settings.yaml eval-grid --split full \
|
| 43 |
+
--checkpoint retention_r50_re10k.ckpt \
|
| 44 |
+
--baselines none --scores r_filter --ratios 0.5 --run-id full_r0.5 --execute
|
| 45 |
+
```
|
| 46 |
+
|
| 47 |
+
Use the same `--ratios` value as the file name (0.25, 0.5 or 0.75).
|
retention_r25_re10k.ckpt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:eafd0413464d62c970ea943f41cd9b8053c76aa75f4eee4633d188f502c92101
|
| 3 |
+
size 2453197055
|
retention_r50_re10k.ckpt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:884cf417cb1c12066d5d63569835bf2b6ac762ca76a6c4df82ef6345492b708d
|
| 3 |
+
size 2453197055
|
retention_r75_re10k.ckpt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b302224cbb0acd81585732a3f7f260afe535c4441401f740dfcc0b03cae2a738
|
| 3 |
+
size 2453197055
|