add

Author: cre185

tools/usage.md

@@ -0,0 +1,74 @@
+# Tools Usage Guide
+
+This document describes utility tools provided with InstantSfM for preprocessing and auxiliary tasks.
+
+## Video Depth Anything
+
+The `video_depth_anything.py` script generates metric depth maps from image sequences using the [Video Depth Anything](https://github.com/DepthAnything/Video-Depth-Anything) model. InstantSfM currently supports only metric depth, so make sure to use the metric depth models. Note that Video Depth Anything requires the input images to be a continuous image sequence (e.g., frames extracted from a video).  
+
+### Setup
+
+You can follow the [official instructions](https://github.com/DepthAnything/Video-Depth-Anything) to set up Video Depth Anything, or follow the steps below:  
+
+**1. Clone Video Depth Anything**
+
+Clone the Video Depth Anything repository into the `external/` directory:
+```bash
+cd external
+git clone https://github.com/DepthAnything/Video-Depth-Anything.git
+cd Video-Depth-Anything
+```
+
+**2. Install Dependencies**
+
+Install the required Python packages:
+```bash
+conda create -n vda python=3.10
+conda activate vda
+pip install -r requirements.txt
+```
+Then install pytorch and xformers as per your CUDA version. For example, for CUDA 12.1:
+```bash
+pip install torch torchvision torchaudio xformers --index-url https://download.pytorch.org/whl/cu121
+```
+
+
+**3. Download Model Checkpoints**
+
+Create a `checkpoints/` directory and download the pretrained weights:
+```bash
+mkdir -p checkpoints
+cd checkpoints
+```
+
+For **metric depth** (recommended):
+```bash
+# Large model (best quality)
+wget https://huggingface.co/depth-anything/Metric-Video-Depth-Anything-Large/resolve/main/metric_video_depth_anything_vitl.pth
+
+# Base model (balanced)
+wget https://huggingface.co/depth-anything/Metric-Video-Depth-Anything-Base/resolve/main/metric_video_depth_anything_vitb.pth
+
+# Small model (fastest)
+wget https://huggingface.co/depth-anything/Metric-Video-Depth-Anything-Small/resolve/main/metric_video_depth_anything_vits.pth
+```
+
+### Usage
+
+**Basic Command**
+
+Process a dataset directory containing images:
+```bash
+python tools/video_depth_anything.py \
+    --data_path /path/to/dataset \
+    --encoder vitl
+```
+
+The script will:
+1. Search for image folders recursively in `data_path`
+2. Process each folder containing images
+3. Save depth maps to `data_path/depth_vda/` by default
+
+**Directory Structure**
+
+The input directory should have exactly the same structure as required by InstantSfM. Use the same `data_path` as for InstantSfM's processing. Output depth maps will be saved in a subdirectory named `depth_vda/`.