Dataset Viewer
The dataset viewer is not available for this dataset.
Job has been terminated due to a temporary spike in resource usage and may be restarted later.
Error code:   JobManagerCrashedError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Hamlyn Dataset — FoundationStereo Depth

数据来源

Hamlyn Centre Laparoscopic / Endoscopic Video Datasets,Kaggle 镜像 mcocoz/hamlyn。本目录为下载后的处理版本,已用 FoundationStereo(NVIDIA,零样本立体匹配 SOTA)重新生成稠密深度图,替换原始深度。


目录结构

hamlyn/
  calibration/
    01/  intrinsics.txt  extrinsics.txt
    04/  ...
    ...
  rectifiedNN/               # 每个序列
    image01/  *.jpg          # 左图(几何左,已校正)
    image02/  *.jpg          # 右图
    depth01/  *.png          # 原始数据集自带深度(参考用)
    depth02/  *.png
    depth01_fs/  *.png       # FoundationStereo 生成的稠密深度(推荐使用)
  _rectifiedNN_fs_vs_gt.png  # 可视化对比图(FS vs 原始 depth)

共 22 个序列:01, 04, 05, 06, 08, 09, 11, 12, 14–27(无 02/03/07/10/13)。


序列概览

序列 分辨率 帧数 FS深度中位 说明
rectified01 640×480 1058 ~81mm
rectified04 360×288 1573 ~41mm
rectified05 360×288 899 ~46mm
rectified06 640×480 1237 ~157mm
rectified08 640×480 14393 ~86mm
rectified09 640×480 806 ~79mm
rectified11 360×288 2427 ~54mm
rectified12 360×288 3389 ~64mm
rectified14 720×288 5742 ~137mm ✅ 宽图有黑边
rectified15 720×288 4155 ~85mm ✅ 宽图有黑边
rectified16 720×288 4169 ~22mm ✅ 宽图有黑边
rectified17 720×288 3651 ~109mm ⚠️ 质量较低
rectified18 720×288 5596 ~40mm ✅ 宽图有黑边
rectified19 720×288 10622 ~123mm ⚠️ 质量较低
rectified20 720×288 2619 ~41mm ✅ 宽图有黑边
rectified21 720×288 5367 ~18mm ✅ 宽图有黑边
rectified22 720×288 2016 ~97mm ✅ 宽图有黑边
rectified23 720×288 11255 ~91mm ⚠️ 质量较低
rectified24 720×288 3674 ~42mm ✅ 宽图有黑边
rectified25 720×288 2121 ~39mm ⚠️ 质量较低
rectified26 720×288 3257 ~54mm ⚠️ 质量较低
rectified27 720×288 2646 ~47mm ✅ 宽图有黑边

⚠️ 质量较低的序列(17/19/23/25/26):FS 深度与原始 depth 尺度偏差较大,可能因过曝帧、弱纹理或标定尺度不一致所致,使用时建议注意。


depth01_fs 编码格式

FoundationStereo 生成的深度统一编码为:

uint16 PNG,0 = 无效像素
深度(mm) = pixel_value / 65535.0 * 250.0

max_depth = 250 mm(覆盖所有序列的有效深度范围)。

解码示例(Python):

import imageio.v2 as iio
import numpy as np

depth_enc = iio.imread("depth01_fs/0000000000.png").astype(np.float32)
depth_mm = depth_enc / 65535.0 * 250.0
depth_mm[depth_enc == 0] = 0  # 保持无效区为0

标定文件格式

每个序列的标定在 calibration/NN/

  • intrinsics.txt:3×4 投影矩阵([fx, 0, cx, 0; 0, fy, cy, 0; 0, 0, 1, 0]
  • extrinsics.txt:左→右相机的 3×4 [R|t],平移 t[0] 为基线(mm)

重要:fx 直接用于 640/360/720 尺寸的图像(不需要按分辨率缩放)。深度公式:

Z (mm) = fx * |t[0]| / disparity_pixels

宽图序列的黑边掩码

rectified14–27(720×288)左右两侧有无效黑边,depth01_fs 中已将以下区域置 0:

  • 左侧:列 0–179
  • 右侧:列 590–719

有效区域为列 180–589(宽 410 像素)。


注意事项

  1. 绝对深度尺度:Hamlyn 标定的 fx 相对图像分辨率系统性偏大(FOV 偏小),各序列的绝对 mm 值可能存在系统误差,不建议跨序列做精确的度量比较。相对深度和几何结构是可靠的。
  2. depth01(原始)vs depth01_fs(FS):原始 depth 有效率仅 19–86%,有大量空洞;FS depth valid 率为 100%,结构更完整平滑。
  3. FS 生成脚本run_hamlyn.py,环境 foundation_stereo,权重 pretrained_models/23-51-11/model_best_bp2.pth
Downloads last month
168