Sync model repo (text/metadata)
Browse files- LICENSE +21 -0
- README.md +171 -0
- SHA256SUMS +10 -0
- benchmarks/zipdepth-base-executorch-fp16-vivo-x300-baseline.yaml +45 -0
- benchmarks/zipdepth-base-executorch-fp16-vivo-x300.yaml +45 -0
- metadata.yaml +23 -0
- samples/im0.jpg +0 -0
- zipdepth_manifest.json +77 -0
LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
MIT License
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2026 Fabio Tosi
|
| 4 |
+
|
| 5 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 6 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 7 |
+
in the Software without restriction, including without limitation the rights
|
| 8 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 9 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 10 |
+
furnished to do so, subject to the following conditions:
|
| 11 |
+
|
| 12 |
+
The above copyright notice and this permission notice shall be included in all
|
| 13 |
+
copies or substantial portions of the Software.
|
| 14 |
+
|
| 15 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 16 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 17 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 18 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 19 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
+
SOFTWARE.
|
README.md
ADDED
|
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: executorch
|
| 3 |
+
license: mit
|
| 4 |
+
base_model: fabiotosi92/ZipDepth
|
| 5 |
+
display_name: ZipDepth Base FP16 — ExecuTorch
|
| 6 |
+
tags:
|
| 7 |
+
- depth-estimation
|
| 8 |
+
- monocular-depth-estimation
|
| 9 |
+
- zipdepth
|
| 10 |
+
- executorch
|
| 11 |
+
- xnnpack
|
| 12 |
+
- fp16
|
| 13 |
+
- arm
|
| 14 |
+
- edge-ai
|
| 15 |
+
pipeline_tag: depth-estimation
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
<!--
|
| 19 |
+
SPDX-FileCopyrightText: Copyright 2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
|
| 20 |
+
|
| 21 |
+
SPDX-License-Identifier: Apache-2.0
|
| 22 |
+
-->
|
| 23 |
+
|
| 24 |
+
# ZipDepth Base FP16 — ExecuTorch
|
| 25 |
+
|
| 26 |
+
This repository packages the unfold-free ZipDepth Base mobile/NPU checkpoint as a static 384×384
|
| 27 |
+
ExecuTorch model. The depth network uses FP16 activations with FP32 serialized convolution weights
|
| 28 |
+
and external tensors.
|
| 29 |
+
|
| 30 |
+
## ✨ Key Highlights
|
| 31 |
+
|
| 32 |
+
- **2.06× faster on Android™ Vivo X300** — one-thread p50 depth-model latency is **20.431 ms**,
|
| 33 |
+
compared with **41.987 ms** for FP32 under the same inference-only protocol.
|
| 34 |
+
- **Negligible quality change** — across all 654 NYUv2 test images, δ1 changes by only
|
| 35 |
+
**−0.046 percentage points** and AbsRel improves by **0.011 percentage points** relative to the
|
| 36 |
+
FP32 depth model under identical fixed-shape preprocessing and evaluation.
|
| 37 |
+
|
| 38 |
+
## 📦 Model Details
|
| 39 |
+
|
| 40 |
+
ZipDepth predicts affine-invariant inverse depth from one RGB image. The package accepts arbitrary
|
| 41 |
+
image dimensions up to 4096×4096, crops the largest top-left square, resizes it to the static
|
| 42 |
+
384×384 depth-model input, and restores the result to the original tensor shape with zero padding.
|
| 43 |
+
|
| 44 |
+
- **Developed by:** Fabio Tosi, Luca Bartolomei, Matteo Poggi, and Stefano Mattoccia
|
| 45 |
+
- **Model type:** Zero-shot monocular relative-depth estimation
|
| 46 |
+
- **License:** MIT
|
| 47 |
+
- **Base model:** [ZipDepth](https://github.com/fabiotosi92/ZipDepth)
|
| 48 |
+
- **Source revision:** [`a302e543`](https://github.com/fabiotosi92/ZipDepth/tree/a302e5437bc58f15c4efd41d3e8222bf24f7d470)
|
| 49 |
+
- **Source checkpoint:** `zipdepth_base_npu.pth`
|
| 50 |
+
- **ExecuTorch revision:** `d13c78971338b35a219e0a27f7045b29c7f4e8c1`
|
| 51 |
+
- **Parameters:** Approximately 6.1 million after inference-time fusion
|
| 52 |
+
- **Artifact size:** 23.47 MiB, 0.27% smaller than the 23.53 MiB FP32 artifact
|
| 53 |
+
|
| 54 |
+
## 🚀 Get Started with the Model
|
| 55 |
+
|
| 56 |
+
### 🔓 Compute Flow — Early Access
|
| 57 |
+
|
| 58 |
+
The inference engine for this model package is available through the Compute Flow Early Access
|
| 59 |
+
Program. Contact ai-early-access@arm.com to request access.
|
| 60 |
+
|
| 61 |
+
## 📊 Quality evaluation
|
| 62 |
+
|
| 63 |
+
Quality was evaluated on all 654 images in the NYUv2 test split. Both models received the same
|
| 64 |
+
top-left square crop resized to 384×384. Their predictions were resized to the ground-truth shape,
|
| 65 |
+
aligned independently in inverse-depth space using a least-squares scale and shift, and evaluated
|
| 66 |
+
over the NYUv2 Eigen crop and 0.001–10 m depth range. This isolates the static depth models; it is
|
| 67 |
+
not an evaluation of the dynamic prepare/restore transforms.
|
| 68 |
+
|
| 69 |
+
| Metric | FP32 | FP16 ExecuTorch | Change |
|
| 70 |
+
| --- | ---: | ---: | ---: |
|
| 71 |
+
| AbsRel ↓ | 15.796% | **15.785%** | 0.011 percentage points lower |
|
| 72 |
+
| δ1 ↑ | 77.055% | 77.009% | 0.046 percentage points lower |
|
| 73 |
+
|
| 74 |
+
The FP16 candidate uses a different official ZipDepth checkpoint because the mobile/NPU checkpoint
|
| 75 |
+
replaces convex unfold upsampling with a deployment-friendly learned blend. The comparison
|
| 76 |
+
therefore measures the complete mobile graph and FP16 conversion, not FP16 rounding in isolation.
|
| 77 |
+
On the distributed 384×384 sample, its output has 0.99645 Pearson correlation with the FP32 output.
|
| 78 |
+
|
| 79 |
+
## 🎯 Performance evaluation
|
| 80 |
+
|
| 81 |
+
The static depth models were profiled using the distributed `samples/im0.jpg` image, which is
|
| 82 |
+
already 384×384. The dynamic prepare and restore models were therefore bypassed.
|
| 83 |
+
|
| 84 |
+
- **CPU:** One Arm CPU thread.
|
| 85 |
+
- **Runs:** One unmeasured in-process inference to initialize each runtime, followed by 5 warmup
|
| 86 |
+
and 30 measured fresh-process runs with 1 second between processes.
|
| 87 |
+
- **Timing scope:** The setup call is not included.
|
| 88 |
+
- **Runtimes:** LiteRT 2.1.6 FP32 baseline and ExecuTorch `d13c7897` FP16 candidate, both with the
|
| 89 |
+
XNNPACK CPU backend.
|
| 90 |
+
- **Target:** Android™ Vivo X300 smartphone (`V2502A`) on Android™ 16, USB powered at 100% battery.
|
| 91 |
+
|
| 92 |
+
| Metric | Android™ Vivo X300: FP32 | Android™ Vivo X300: FP16 | Uplift |
|
| 93 |
+
| --- | ---: | ---: | ---: |
|
| 94 |
+
| p50 inference latency | 41.987 ms | **20.431 ms** | **2.06× faster** |
|
| 95 |
+
| p90 inference latency | 42.612 ms | **21.342 ms** | **2.00× faster** |
|
| 96 |
+
| p99 inference latency | 43.236 ms | **27.181 ms** | **1.59× faster** |
|
| 97 |
+
|
| 98 |
+
## 🛠️ Technical Specifications
|
| 99 |
+
|
| 100 |
+
### Runtime Architecture
|
| 101 |
+
|
| 102 |
+
| Component role | Framework / format |
|
| 103 |
+
| --- | --- |
|
| 104 |
+
| Input square crop and resize | ExecuTorch / PTE |
|
| 105 |
+
| ZipDepth Base inference | ExecuTorch / PTE |
|
| 106 |
+
| Output resize and zero padding | ExecuTorch / PTE |
|
| 107 |
+
|
| 108 |
+
### Precision
|
| 109 |
+
|
| 110 |
+
- Internal depth-model activations use FP16.
|
| 111 |
+
- Public input and output tensors remain FP32 to preserve the existing brick contract.
|
| 112 |
+
- Dynamic image-shape transforms use FP32.
|
| 113 |
+
- The fixed 24×24-to-12×12 adaptive average pool is represented as an equivalent 2×2,
|
| 114 |
+
stride-2 average pool supported by the deployed ExecuTorch kernel set.
|
| 115 |
+
|
| 116 |
+
### Input Specification
|
| 117 |
+
|
| 118 |
+
| Input | Data type | Shape | Value range |
|
| 119 |
+
| --- | --- | --- | --- |
|
| 120 |
+
| RGB image | FP32 | `3 × H × W`, with `1 ≤ H,W ≤ 4096` | `[0, 1]` |
|
| 121 |
+
|
| 122 |
+
### Output Specification
|
| 123 |
+
|
| 124 |
+
| Output | Data type | Shape | Description |
|
| 125 |
+
| --- | --- | --- | --- |
|
| 126 |
+
| Relative depth | FP32 | `H × W` | Affine-invariant inverse depth; padded regions are zero |
|
| 127 |
+
|
| 128 |
+
### Repository Contents
|
| 129 |
+
|
| 130 |
+
- `zipdepth_base_fp16.pte` — static FP16 ZipDepth Base mobile/NPU depth model.
|
| 131 |
+
- `zipdepth_prepare_image.pte` — dynamic input crop-and-resize transform.
|
| 132 |
+
- `zipdepth_restore_depth.pte` — dynamic output resize-and-padding transform.
|
| 133 |
+
- `zipdepth_manifest.json` — model tensor contracts and runtime configuration.
|
| 134 |
+
- `samples/im0.jpg` — 384×384 upstream sample used for profiling.
|
| 135 |
+
- `metadata.yaml` and `benchmarks/` — model and benchmark metadata.
|
| 136 |
+
- `LICENSE` — upstream MIT license.
|
| 137 |
+
- `SHA256SUMS` — reproducibility checksums.
|
| 138 |
+
|
| 139 |
+
## ⚠️ Known Limitations
|
| 140 |
+
|
| 141 |
+
- The square-crop package pipeline does not preserve the aspect-ratio preprocessing used for
|
| 142 |
+
ZipDepth's published benchmark results.
|
| 143 |
+
- The output is affine-invariant relative depth, not metric depth.
|
| 144 |
+
- The model has no temporal-consistency mechanism and may flicker when applied independently to
|
| 145 |
+
video frames.
|
| 146 |
+
|
| 147 |
+
## 🗂️ Model and Asset Origin
|
| 148 |
+
|
| 149 |
+
- [`zipdepth_base_npu.pth`](https://github.com/fabiotosi92/ZipDepth/tree/a302e5437bc58f15c4efd41d3e8222bf24f7d470/checkpoints)
|
| 150 |
+
is the source checkpoint for `zipdepth_base_fp16.pte`.
|
| 151 |
+
- [`samples/im0.jpg`](https://github.com/fabiotosi92/ZipDepth/blob/a302e5437bc58f15c4efd41d3e8222bf24f7d470/assets/examples/im0.jpg)
|
| 152 |
+
is an upstream example crop.
|
| 153 |
+
|
| 154 |
+
## 🔐 Checksums
|
| 155 |
+
|
| 156 |
+
Verify the checked-out files with:
|
| 157 |
+
|
| 158 |
+
```sh
|
| 159 |
+
shasum -a 256 -c SHA256SUMS
|
| 160 |
+
```
|
| 161 |
+
|
| 162 |
+
## Citation
|
| 163 |
+
|
| 164 |
+
```bibtex
|
| 165 |
+
@inproceedings{tosi2026zipdepth,
|
| 166 |
+
title = {{ZipDepth}: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device},
|
| 167 |
+
author = {Tosi, Fabio and Bartolomei, Luca and Poggi, Matteo and Mattoccia, Stefano},
|
| 168 |
+
booktitle = {European Conference on Computer Vision (ECCV)},
|
| 169 |
+
year = {2026}
|
| 170 |
+
}
|
| 171 |
+
```
|
SHA256SUMS
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0007e2ff761f1b89ad89327870b807cd4de00cb657b442b62de2f92fdc87d508 LICENSE
|
| 2 |
+
3e2361de413265dcef61f7758a29297776d8ecaa72d7041917f033d46c05cba0 README.md
|
| 3 |
+
0ae0b8c374363c6b82434632143b1c0dfbf171d0cdf779f71c597ceeb3abd34e benchmarks/zipdepth-base-executorch-fp16-vivo-x300-baseline.yaml
|
| 4 |
+
47ad0825549f7e8fa6b4424e595a801437cf3d1be53ee37e3beaf6d18083f670 benchmarks/zipdepth-base-executorch-fp16-vivo-x300.yaml
|
| 5 |
+
47f267cad4bb732b7672c8237e6de1ecd7d53b24259c280dac7bc929fa4744f7 metadata.yaml
|
| 6 |
+
ac960f4f6fecc05ce47411ad030974510138e209aef370662aa4d81fc26decde samples/im0.jpg
|
| 7 |
+
33f72f8c10ae5449872816e77f7e84f055a35c55588fbfbd1b41dd633e146a6f zipdepth_base_fp16.pte
|
| 8 |
+
16b6bec7550ed15d10ede2dd456e8d7b7fc61d07019eddf5cf0d25fd2219cf42 zipdepth_manifest.json
|
| 9 |
+
b59e2cf4db36efa2d7cdfd3695f000bd8e1b36bf551f5eb4bc0971db505d5997 zipdepth_prepare_image.pte
|
| 10 |
+
1e37a26736d90ac80a269e77f8a58e232cdb5311f4c504bc387828866b3af78b zipdepth_restore_depth.pte
|
benchmarks/zipdepth-base-executorch-fp16-vivo-x300-baseline.yaml
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-FileCopyrightText: Copyright 2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
|
| 2 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 3 |
+
|
| 4 |
+
version: 1.2.0
|
| 5 |
+
report_type: image-depth-estimation
|
| 6 |
+
profile: Baseline
|
| 7 |
+
created_at: '2026-09-17T11:11:10Z'
|
| 8 |
+
context:
|
| 9 |
+
target_ref: vivo-x300
|
| 10 |
+
runtime:
|
| 11 |
+
name: litert
|
| 12 |
+
execution_backend: cpu
|
| 13 |
+
config:
|
| 14 |
+
n_threads: 1
|
| 15 |
+
optimisations:
|
| 16 |
+
- XNNPACK
|
| 17 |
+
version: 2.1.6
|
| 18 |
+
dataset:
|
| 19 |
+
name: NYU Depth V2
|
| 20 |
+
version: V2
|
| 21 |
+
slice: labeled test split (Eigen crop)
|
| 22 |
+
sample_count: 654
|
| 23 |
+
reference_url: https://cs.nyu.edu/~fergus/datasets/nyu_depth_v2.html
|
| 24 |
+
benchmark:
|
| 25 |
+
batch_size: 1
|
| 26 |
+
image_width: 384
|
| 27 |
+
image_height: 384
|
| 28 |
+
num_runs: 30
|
| 29 |
+
warmup_runs: 5
|
| 30 |
+
mode: offline
|
| 31 |
+
pipeline_scope: depth-model-only
|
| 32 |
+
performance:
|
| 33 |
+
end_to_end_latency_ms:
|
| 34 |
+
p50: 41.987
|
| 35 |
+
p90: 42.612
|
| 36 |
+
p99: 43.236
|
| 37 |
+
throughput_images_per_second: 23.8167
|
| 38 |
+
accuracy:
|
| 39 |
+
delta1: 0.770555
|
| 40 |
+
absolute_relative_error: 0.157965
|
| 41 |
+
alignment: per-image least-squares scale and shift in disparity space
|
| 42 |
+
valid_depth_range_m:
|
| 43 |
+
minimum: 0.001
|
| 44 |
+
maximum: 10.0
|
| 45 |
+
sample_count: 654
|
benchmarks/zipdepth-base-executorch-fp16-vivo-x300.yaml
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-FileCopyrightText: Copyright 2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
|
| 2 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 3 |
+
|
| 4 |
+
version: 1.2.0
|
| 5 |
+
report_type: image-depth-estimation
|
| 6 |
+
profile: Arm-Optimized
|
| 7 |
+
created_at: '2026-09-17T11:10:17Z'
|
| 8 |
+
context:
|
| 9 |
+
target_ref: vivo-x300
|
| 10 |
+
runtime:
|
| 11 |
+
name: executorch
|
| 12 |
+
execution_backend: cpu
|
| 13 |
+
config:
|
| 14 |
+
n_threads: 1
|
| 15 |
+
optimisations:
|
| 16 |
+
- XNNPACK
|
| 17 |
+
version: d13c78971338b35a219e0a27f7045b29c7f4e8c1
|
| 18 |
+
dataset:
|
| 19 |
+
name: NYU Depth V2
|
| 20 |
+
version: V2
|
| 21 |
+
slice: labeled test split (Eigen crop)
|
| 22 |
+
sample_count: 654
|
| 23 |
+
reference_url: https://cs.nyu.edu/~fergus/datasets/nyu_depth_v2.html
|
| 24 |
+
benchmark:
|
| 25 |
+
batch_size: 1
|
| 26 |
+
image_width: 384
|
| 27 |
+
image_height: 384
|
| 28 |
+
num_runs: 30
|
| 29 |
+
warmup_runs: 5
|
| 30 |
+
mode: offline
|
| 31 |
+
pipeline_scope: depth-model-only
|
| 32 |
+
performance:
|
| 33 |
+
end_to_end_latency_ms:
|
| 34 |
+
p50: 20.431
|
| 35 |
+
p90: 21.342
|
| 36 |
+
p99: 27.181
|
| 37 |
+
throughput_images_per_second: 48.9444
|
| 38 |
+
accuracy:
|
| 39 |
+
delta1: 0.770092
|
| 40 |
+
absolute_relative_error: 0.157853
|
| 41 |
+
alignment: per-image least-squares scale and shift in disparity space
|
| 42 |
+
valid_depth_range_m:
|
| 43 |
+
minimum: 0.001
|
| 44 |
+
maximum: 10.0
|
| 45 |
+
sample_count: 654
|
metadata.yaml
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# SPDX-FileCopyrightText: Copyright 2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
|
| 2 |
+
# SPDX-License-Identifier: Apache-2.0
|
| 3 |
+
|
| 4 |
+
schema_version: 1.0.0
|
| 5 |
+
report_type: image-depth-estimation
|
| 6 |
+
task_type: monocular-depth-estimation
|
| 7 |
+
title: ZipDepth FP16 ExecuTorch — Compute Flow
|
| 8 |
+
description: >-
|
| 9 |
+
Monocular relative-depth package derived from the unfold-free ZipDepth Base mobile/NPU
|
| 10 |
+
checkpoint. The package uses a static ExecuTorch depth model with FP32 tensor interfaces,
|
| 11 |
+
FP16 activations, FP32 serialized convolution weights, and dynamic ExecuTorch transforms to
|
| 12 |
+
preserve the source image shape at its boundaries.
|
| 13 |
+
family_name: ZipDepth
|
| 14 |
+
id: Arm/zipdepth-fp16-executorch
|
| 15 |
+
filename: zipdepth_base_fp16.pte
|
| 16 |
+
base_model_id: fabiotosi92/ZipDepth
|
| 17 |
+
vendor: University of Bologna
|
| 18 |
+
base_model_url: https://github.com/fabiotosi92/ZipDepth
|
| 19 |
+
weight_dtype: fp32
|
| 20 |
+
profile: Arm-Optimized
|
| 21 |
+
model_size_mb: 24.609984
|
| 22 |
+
parameter_count: 6142058
|
| 23 |
+
format: executorch
|
samples/im0.jpg
ADDED
|
zipdepth_manifest.json
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"brick": {
|
| 3 |
+
"type": "monocular_depth_estimation",
|
| 4 |
+
"name": "ZipDepth",
|
| 5 |
+
"version": "1.0"
|
| 6 |
+
},
|
| 7 |
+
"models": {
|
| 8 |
+
"resize_input": {
|
| 9 |
+
"runner": "executorch",
|
| 10 |
+
"filename": "zipdepth_prepare_image.pte",
|
| 11 |
+
"inputs": {
|
| 12 |
+
"image": {
|
| 13 |
+
"index": 0,
|
| 14 |
+
"dtype": "float32",
|
| 15 |
+
"shape": [1, 3, -1, -1]
|
| 16 |
+
}
|
| 17 |
+
},
|
| 18 |
+
"outputs": {
|
| 19 |
+
"image": {
|
| 20 |
+
"index": 0,
|
| 21 |
+
"dtype": "float32",
|
| 22 |
+
"shape": [1, 3, 384, 384]
|
| 23 |
+
}
|
| 24 |
+
},
|
| 25 |
+
"runner_params": {}
|
| 26 |
+
},
|
| 27 |
+
"zip_depth": {
|
| 28 |
+
"runner": "executorch",
|
| 29 |
+
"filename": "zipdepth_base_fp16.pte",
|
| 30 |
+
"inputs": {
|
| 31 |
+
"image": {
|
| 32 |
+
"index": 0,
|
| 33 |
+
"dtype": "float32",
|
| 34 |
+
"shape": [1, 3, 384, 384]
|
| 35 |
+
}
|
| 36 |
+
},
|
| 37 |
+
"outputs": {
|
| 38 |
+
"relative_depth": {
|
| 39 |
+
"index": 0,
|
| 40 |
+
"dtype": "float32",
|
| 41 |
+
"shape": [1, 1, 384, 384]
|
| 42 |
+
}
|
| 43 |
+
},
|
| 44 |
+
"runner_params": {}
|
| 45 |
+
},
|
| 46 |
+
"resize_output": {
|
| 47 |
+
"runner": "executorch",
|
| 48 |
+
"filename": "zipdepth_restore_depth.pte",
|
| 49 |
+
"inputs": {
|
| 50 |
+
"relative_depth": {
|
| 51 |
+
"index": 0,
|
| 52 |
+
"dtype": "float32",
|
| 53 |
+
"shape": [1, 1, 384, 384]
|
| 54 |
+
},
|
| 55 |
+
"image": {
|
| 56 |
+
"index": 1,
|
| 57 |
+
"dtype": "float32",
|
| 58 |
+
"shape": [1, 3, -1, -1]
|
| 59 |
+
}
|
| 60 |
+
},
|
| 61 |
+
"outputs": {
|
| 62 |
+
"relative_depth": {
|
| 63 |
+
"index": 0,
|
| 64 |
+
"dtype": "float32",
|
| 65 |
+
"shape": [1, 1, -1, -1]
|
| 66 |
+
}
|
| 67 |
+
},
|
| 68 |
+
"runner_params": {}
|
| 69 |
+
}
|
| 70 |
+
},
|
| 71 |
+
"constants": {
|
| 72 |
+
"maximum_dynamic_dimension": {
|
| 73 |
+
"type": "int64",
|
| 74 |
+
"value": 4096
|
| 75 |
+
}
|
| 76 |
+
}
|
| 77 |
+
}
|