Instructions to use wardner/Depth-Anything-V3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- RKLLM
How to use wardner/Depth-Anything-V3 with RKLLM:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
File size: 1,943 Bytes
b8ed715 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 | {
"label": "[Depth]视觉深度信息v3",
"capability_types": [
"swcap_vision_depth"
],
"spec_factory": "depth_anything_v3",
"worker_name": "depth-v3",
"model_dir_name": "Depth-Anything-V3",
"model_dir_env": "WDNAI_DEPTH_ANYTHING_V3_DIR",
"instance_name": "depth-v3",
"memory_estimate_mb": 3000,
"estimated_init_seconds": 30,
"description": "视觉深度信息v3,给定图像输出逐像素深度图与置信度(Depth Anything V3 单目深度估计)",
"notes": "【当前部署】1-view / 280×280 导出(global 段输入 [1,401,1,768]、head 段输出 [1,1,280,280])—— 一张图即可出深度图 + 置信度。view 数是**导出时**参数(global 段输入形状 [V,L,1,D],不是运行期参数),换 V 只需重新导出三段 RKNN(model-zoo 的 --views,已验证范围 1-10),**SDK 与本服务无需改动**。部署侧文件命名跟 LocateAnything-3B 一致:<role>_DepthAnythingV3.280x280.1view.1828.rknn + 同名 .weight(去掉 .1828 段)—— 共 6 个文件。多相机几何另需 model-zoo 导出 ray 分支(当前导出脚本只出 depth+depth_conf,不支持)并换更大的模型——论文数据显示 DA3-Base 的位姿精度仅为 DA3-Giant 的 1/4,位姿比深度更吃模型规模。单目深度估计本身(kind=\"depth\",图像 → 深度图 + 置信度,不是文本)。RKNN 导出被拆成三段:local(逐图 token)→ global(跨 view 融合)→ head(出深度/置信度),三段**共享**一份按 NPU 核分配的内部显存(C++ demo 的 da3_init_internal_mem_share,约 1.6GB),少了这一步 model_init 会 MODEL_SETUP fail。权重目录由 wdnai 内部的 _default_depth_anything_v3_paths 解析,可用 WDNAI_DEPTH_ANYTHING_V3_DIR 覆写。模型未部署时 load 会抛 FileNotFoundError,由 worker 的 load 流程转成 errored 状态。",
"available_targets": [
"182x"
],
"display_order": 80
}
|