depthpro_cvb.npz dict | sam2_cvb.npz dict | vggt_cvb_0.npz dict | vggt_cvb_1.npz dict | vggt_cvb_2.npz dict | vggt_cvb_3.npz dict | vggt_cvb_4.npz dict | vggt_cvb_5.npz dict |
|---|---|---|---|---|---|---|---|
{
"bytes": 132579804,
"md5": "94e7f1c32a56c6d49912ef5237488a3a",
"sha256": "9c2bf7681704199b39784a666a15852eb4e878185886b97ad538c562743d84d6"
} | {
"bytes": 755467,
"md5": "9a11ba37a21af2b3e44df3d790913d4b",
"sha256": "75b03b8463fe547e03197ea109c1950022008d86dd60c22fc97c40b51f6c8072"
} | {
"bytes": 27047098,
"md5": "795a5122085558d3d8776d0f75d41df6",
"sha256": "4d48768078d00acb3aec828d942d0b61af6d0750c59450c7525be0d332320ee3"
} | {
"bytes": 26716909,
"md5": "ab7983762e94c65e672da58682c8f279",
"sha256": "4c2a1e4d5b61358844de4bed743c9861958b73d38692ec734c1c34540f46ed8b"
} | {
"bytes": 28727464,
"md5": "73de8370426a85ac0fdb0da1bbf34d25",
"sha256": "d70e6fa742eb1475140436065838d42ad431bf2a05b0df3c19d86204317593c9"
} | {
"bytes": 28423040,
"md5": "9178703313662657f24005a7cbdd5edd",
"sha256": "e586e1569332a84949b7bd0ec7f635bf5792ddd0cb926be20900a7ff0858876a"
} | {
"bytes": 22551378,
"md5": "bb174320dc6589d62cad0575ccecf824",
"sha256": "11c5da098b9accb38d991ab77d63e69cd26037719039484e913d9fd6786e11d8"
} | {
"bytes": 23127522,
"md5": "e15b568304967fc3851c921bdb20b615",
"sha256": "79d9feba4bf12a1241087c7f380850c413d74c1f98f25f32a7f4f5b01388e42e"
} |
Off-the-shelf model outputs on the 600 CV-Bench depth pairs
Raw outputs of three public vision models on the 600 items of the CV-Bench 3D Depth task
(nyu-visionx/CV-Bench, test_3d.jsonl, the items with
task == "Depth" in file order). The item number i used below is the position of an item among these 600 (0-599).
No images are included; get them from CV-Bench.
| File | Model | Content |
|---|---|---|
depthpro_cvb.npz |
DepthPro, ONNX export with q4 weights (model_q4.onnx, file of 2024-10-09) |
idx (600 item numbers), pred (600 x 256 x 256 float32, the raw first output of the model for the image resized to 256 x 256 with bilinear filtering and ImageNet normalisation), focal (600, second model output), size (600 x 2, original image width and height) |
sam2_cvb.npz |
SAM 2, hiera-small | one mask per ground-truth box (two boxes per item, in the order of the item's bbox field), prompted with the box (coordinates truncated to integers), single-mask output. idx (600), shape (1,200 x 2, mask height and width), score (1,200, predicted IoU), bits and off: mask j is np.unpackbits(bits[off[j]:off[j+1]])[:h*w].reshape(h, w) |
vggt_cvb_0.npz ... vggt_cvb_5.npz |
VGGT-1B, single image | for item i in chunk i // 100: d{i} depth map (float32), c{i} depth confidence (float16), E{i} extrinsics (3 x 4, identity for the single frame), K{i} intrinsics (3 x 3); the image is re-encoded as JPEG and preprocessed with VGGT's load_and_preprocess_images |
checksums.json lists the size, sha256 and md5 of every file.
import numpy as np
z = np.load("sam2_cvb.npz")
j = 0 # first box of the first item
h, w = z["shape"][j]
mask = np.unpackbits(z["bits"][z["off"][j]:z["off"][j + 1]])[: h * w].reshape(h, w).astype(bool)
Licence
These are outputs of third-party models on third-party images. Use them under the terms of CV-Bench and of the respective models (DepthPro, SAM 2, VGGT-1B; see each model's licence).
- Downloads last month
- 38