MiDaS v2.1 | Luxonis

Our new model ZOO works with DepthAI V3. Find out more in our documentation.

0 Likes

Model Details

Model Description

MiDaS v2.1 is a robust model for monocular depth estimation. It calculates the relative distance of objects from the camera and is reported to perform well across various scenarios. We implement here the Small version of the model, which is optimized for use on edge devices.

Developed by: René Ranftl et al.
Shared by:
Model type: Computer Vision
License:
Resources for more information:

Photo by James Ranieri from

Training Details

Training Data

The model was trained on a diverse set of data taken from the ReDWeb, DIML, Movies, MegaDepth, WSVD, TartanAir, ApolloScape, BlendedMVS, and IRS datasets. Training on depth data of different modalities was enabled by multi-objective optimization.

Testing Details

Metrics

Test metrics were evaluated on various previously unseen datasets to demonstrate the model's generalizability:

DIW dataset test split,
ETH3D dataset where ground truth is available,
Sintel dataset where ground truth is available,
KITTI dataset validation split for depth estimation and the Eigen test split,
NYU dataset test split,
TUM dataset subset of humans in indoor environments.

The authors used metrics aliged with the dataset ground truth:

Weighted Human Disagreement Rate (WHDR) For DIW,
Mean Absolute Relative Error (AbsRel) for ETH3D and Sintel,
Percentage of pixels with δ>1.25 for KITTI, NYU, and TUM.

Dataset	WHDR	AbsRel	δ>1.25
DIW	0.1344	/	/
ETH3D	/	0.1344	/
Sintel	/	0.3370	/
Kitti	/	/	29.27
NYU	/	/	13.43
TUM	/	/	14.53

The results are taken from the page on GitHub.

Technical Specifications

Input/Output Details

Input:
- Name: image
  - Info: NCHW BGR image
Output:
- Name: relative_depth
  - Info: 2D map of relative depth values

Model Architecture

Backbone: EfficientNet Lite 3

For more information, see the source code.

Throughput

Model variant: midas-v2-1:small-384x256

• Input shape: [1, 3, 256, 384] • Output shape: [1, 256, 384]

• Params (M): 16.564 • GFLOPs: 7.001

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	9.89	N/A
RVC4	INT8	461.41	5.10

Model variant: midas-v2-1:small-256x192

• Input shape: [1, 3, 192, 256] • Output shape: [1, 192, 256]

• Params (M): 16.564 • GFLOPs: 3.501

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	20.24	N/A
RVC4	INT8	616.54	4.43

Model variant: midas-v2-1:small-512x288

• Input shape: [1, 3, 288, 512] • Output shape: [1, 288, 512]

• Params (M): 16.564 • GFLOPs: 10.502

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	7.08	N/A
RVC4	FP32	106.53	3.28

Model variant: midas-v2-1:small-512x384

• Input shape: [1, 3, 384, 512] • Output shape: [1, 384, 512]

• Params (M): 16.564 • GFLOPs: 14.002

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	4.80	N/A
RVC4	INT8	295.30	4.70

Model variant: midas-v2-1:small-1024x768

• Input shape: [1, 3, 768, 1024] • Output shape: [1, 768, 1024]

• Params (M): 16.564 • GFLOPs: 56.009

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC4	INT8	70.59	4.78

* Benchmarked with , using 2 threads (and the DSP runtime in balanced mode for RVC4).

* Parameters and FLOPs are obtained from the package.

Quantization

RVC4 models were quantized to int8 using the HubAI General dataset.

Utilization

Models converted for RVC Platforms can be used for inference on OAK devices. DepthAI pipelines are used to define the information flow linking the device, inference model, and the output parser (as defined in model head(s)). Below, we present the most crucial utilization steps for the particular model. Please consult the docs for more information.

Install DAIv3 and depthai-nodes libraries:

pip install depthai
pip install depthai-nodes

Define model:

model_description = dai.NNModelDescription(
    "luxonis/midas-v2-1:small-384x256"
)

nn = pipeline.create(ParsingNeuralNetwork).build(
    <CameraNode>, model_description
)

Inspect model head(s):

MapOutputParser that outputs message (depth map).

Get parsed output(s):

while pipeline.isRuning():
    parser_output: Map2D = parser_output_queue.get()

Example

You can quickly run the model using our script. It automatically downloads the model, creates a DepthAI pipeline, runs the inference, and displays the results using our DepthAI visualizer tool. To try it out, run:

python3 main.py \
    --model luxonis/midas-v2-1:small-384x256 \
    -overlay

Monocular depth estimation model.
License	MIT Commercial use
Downloads	2088
Tasks	Depth Estimation
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC4	9 months ago
		RVC2, RVC4	9 months ago
		RVC2, RVC4	9 months ago
		RVC2, RVC4	9 months ago
		RVC2, RVC4	9 months ago