3D-Deep-BOX: Optimized for Mobile Deployment

Real-time 3D object detection

3D Deep Box is a machine learning model that predicts 3D bounding boxes and classes of objects in an image.

This model is an implementation of 3D-Deep-BOX found here.

This repository provides scripts to run 3D-Deep-BOX on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

Model Type: Object detection
Model Stats:
- Model checkpoint: YOLOv3-tiny
- Input resolution(YOLO): 224x640
- Number of parameters(YOLO): 8.85M
- Model size(YOLO): 37.3 MB
- Input resolution(VGG): 224x224
- Number of parameters(VGG): 144M
- Model size(VGG): 175.9 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
Yolo	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	21.944 ms	0 - 135 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	2.435 ms	2 - 4 MB	FP16	NPU	3D-Deep-BOX.so
Yolo	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	5.173 ms	3 - 54 MB	FP16	NPU	3D-Deep-BOX.onnx
Yolo	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	16.556 ms	0 - 61 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	1.748 ms	2 - 22 MB	FP16	NPU	3D-Deep-BOX.so
Yolo	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	3.528 ms	1 - 26 MB	FP16	NPU	3D-Deep-BOX.onnx
Yolo	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	14.082 ms	0 - 33 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	2.087 ms	2 - 19 MB	FP16	NPU	Use Export Script
Yolo	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	4.32 ms	2 - 15 MB	FP16	NPU	3D-Deep-BOX.onnx
Yolo	SA7255P ADP	SA7255P	TFLITE	67.784 ms	0 - 29 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	SA7255P ADP	SA7255P	QNN	34.599 ms	2 - 11 MB	FP16	NPU	Use Export Script
Yolo	SA8255 (Proxy)	SA8255P Proxy	TFLITE	22.109 ms	0 - 141 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	SA8255 (Proxy)	SA8255P Proxy	QNN	2.436 ms	2 - 4 MB	FP16	NPU	Use Export Script
Yolo	SA8295P ADP	SA8295P	TFLITE	23.94 ms	0 - 31 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	SA8295P ADP	SA8295P	QNN	3.475 ms	0 - 18 MB	FP16	NPU	Use Export Script
Yolo	SA8650 (Proxy)	SA8650P Proxy	TFLITE	21.84 ms	0 - 140 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	SA8650 (Proxy)	SA8650P Proxy	QNN	2.426 ms	2 - 4 MB	FP16	NPU	Use Export Script
Yolo	SA8775P ADP	SA8775P	TFLITE	27.669 ms	0 - 29 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	SA8775P ADP	SA8775P	QNN	3.837 ms	2 - 11 MB	FP16	NPU	Use Export Script
Yolo	QCS8275 (Proxy)	QCS8275 Proxy	TFLITE	67.784 ms	0 - 29 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	QCS8275 (Proxy)	QCS8275 Proxy	QNN	34.599 ms	2 - 11 MB	FP16	NPU	Use Export Script
Yolo	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	21.809 ms	0 - 136 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	QCS8550 (Proxy)	QCS8550 Proxy	QNN	2.44 ms	2 - 5 MB	FP16	NPU	Use Export Script
Yolo	QCS9075 (Proxy)	QCS9075 Proxy	TFLITE	27.669 ms	0 - 29 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	QCS9075 (Proxy)	QCS9075 Proxy	QNN	3.837 ms	2 - 11 MB	FP16	NPU	Use Export Script
Yolo	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	22.505 ms	0 - 60 MB	FP16	NPU	3D-Deep-BOX.tflite
Yolo	QCS8450 (Proxy)	QCS8450 Proxy	QNN	3.531 ms	2 - 22 MB	FP16	NPU	Use Export Script
Yolo	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	2.653 ms	2 - 2 MB	FP16	NPU	Use Export Script
Yolo	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	5.558 ms	11 - 11 MB	FP16	NPU	3D-Deep-BOX.onnx
VGG	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	4.72 ms	0 - 667 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	4.775 ms	1 - 2 MB	FP16	NPU	3D-Deep-BOX.so
VGG	Samsung Galaxy S23	Snapdragon® 8 Gen 2	ONNX	5.323 ms	0 - 511 MB	FP16	NPU	3D-Deep-BOX.onnx
VGG	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	3.572 ms	0 - 125 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	3.753 ms	0 - 18 MB	FP16	NPU	3D-Deep-BOX.so
VGG	Samsung Galaxy S24	Snapdragon® 8 Gen 3	ONNX	4.057 ms	0 - 82 MB	FP16	NPU	3D-Deep-BOX.onnx
VGG	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	3.454 ms	0 - 78 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	3.427 ms	1 - 76 MB	FP16	NPU	Use Export Script
VGG	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	ONNX	3.811 ms	1 - 75 MB	FP16	NPU	3D-Deep-BOX.onnx
VGG	SA7255P ADP	SA7255P	TFLITE	257.893 ms	0 - 73 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	SA7255P ADP	SA7255P	QNN	257.859 ms	1 - 10 MB	FP16	NPU	Use Export Script
VGG	SA8255 (Proxy)	SA8255P Proxy	TFLITE	4.721 ms	0 - 669 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	SA8255 (Proxy)	SA8255P Proxy	QNN	4.801 ms	1 - 3 MB	FP16	NPU	Use Export Script
VGG	SA8295P ADP	SA8295P	TFLITE	9.753 ms	0 - 76 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	SA8295P ADP	SA8295P	QNN	9.913 ms	1 - 19 MB	FP16	NPU	Use Export Script
VGG	SA8650 (Proxy)	SA8650P Proxy	TFLITE	4.729 ms	0 - 664 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	SA8650 (Proxy)	SA8650P Proxy	QNN	4.776 ms	1 - 2 MB	FP16	NPU	Use Export Script
VGG	SA8775P ADP	SA8775P	TFLITE	10.811 ms	0 - 73 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	SA8775P ADP	SA8775P	QNN	10.713 ms	0 - 10 MB	FP16	NPU	Use Export Script
VGG	QCS8275 (Proxy)	QCS8275 Proxy	TFLITE	257.893 ms	0 - 73 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	QCS8275 (Proxy)	QCS8275 Proxy	QNN	257.859 ms	1 - 10 MB	FP16	NPU	Use Export Script
VGG	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	4.721 ms	0 - 668 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	QCS8550 (Proxy)	QCS8550 Proxy	QNN	4.772 ms	1 - 4 MB	FP16	NPU	Use Export Script
VGG	QCS9075 (Proxy)	QCS9075 Proxy	TFLITE	10.811 ms	0 - 73 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	QCS9075 (Proxy)	QCS9075 Proxy	QNN	10.713 ms	0 - 10 MB	FP16	NPU	Use Export Script
VGG	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	8.44 ms	0 - 125 MB	FP16	NPU	3D-Deep-BOX.tflite
VGG	QCS8450 (Proxy)	QCS8450 Proxy	QNN	8.649 ms	1 - 83 MB	FP16	NPU	Use Export Script
VGG	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	5.023 ms	1 - 1 MB	FP16	NPU	Use Export Script
VGG	Snapdragon X Elite CRD	Snapdragon® X Elite	ONNX	5.251 ms	89 - 89 MB	FP16	NPU	3D-Deep-BOX.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[deepbox]"

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.deepbox.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.deepbox.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

Performance check on-device on a cloud-hosted device
Downloads compiled assets that can be deployed on-device for Android.
Accuracy check between PyTorch and on-device outputs.

python -m qai_hub_models.models.deepbox.export

Profiling Results
------------------------------------------------------------
Yolo
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 21.9                   
Estimated peak memory usage (MB): [0, 135]               
Total # Ops                     : 128                    
Compute Unit(s)                 : NPU (128 ops)          

------------------------------------------------------------
VGG
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 4.7                    
Estimated peak memory usage (MB): [0, 667]               
Total # Ops                     : 40                     
Compute Unit(s)                 : NPU (40 ops)

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.deepbox import Model

# Load the model
model = Model.from_pretrained()
bbox2D_dectector_model = model.bbox2D_dectector
bbox3D_dectector_model = model.bbox3D_dectector

# Device
device = hub.Device("Samsung Galaxy S23")

# Trace model
bbox2D_dectector_input_shape = bbox2D_dectector_model.get_input_spec()
bbox2D_dectector_sample_inputs = bbox2D_dectector_model.sample_inputs()

traced_bbox2D_dectector_model = torch.jit.trace(bbox2D_dectector_model, [torch.tensor(data[0]) for _, data in bbox2D_dectector_sample_inputs.items()])

# Compile model on a specific device
bbox2D_dectector_compile_job = hub.submit_compile_job(
    model=traced_bbox2D_dectector_model ,
    device=device,
    input_specs=bbox2D_dectector_model.get_input_spec(),
)

# Get target model to run on-device
bbox2D_dectector_target_model = bbox2D_dectector_compile_job.get_target_model()
# Trace model
bbox3D_dectector_input_shape = bbox3D_dectector_model.get_input_spec()
bbox3D_dectector_sample_inputs = bbox3D_dectector_model.sample_inputs()

traced_bbox3D_dectector_model = torch.jit.trace(bbox3D_dectector_model, [torch.tensor(data[0]) for _, data in bbox3D_dectector_sample_inputs.items()])

# Compile model on a specific device
bbox3D_dectector_compile_job = hub.submit_compile_job(
    model=traced_bbox3D_dectector_model ,
    device=device,
    input_specs=bbox3D_dectector_model.get_input_spec(),
)

# Get target model to run on-device
bbox3D_dectector_target_model = bbox3D_dectector_compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

bbox2D_dectector_profile_job = hub.submit_profile_job(
    model=bbox2D_dectector_target_model,
    device=device,
)
bbox3D_dectector_profile_job = hub.submit_profile_job(
    model=bbox3D_dectector_target_model,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

bbox2D_dectector_input_data = bbox2D_dectector_model.sample_inputs()
bbox2D_dectector_inference_job = hub.submit_inference_job(
    model=bbox2D_dectector_target_model,
    device=device,
    inputs=bbox2D_dectector_input_data,
)
bbox2D_dectector_inference_job.download_output_data()
bbox3D_dectector_input_data = bbox3D_dectector_model.sample_inputs()
bbox3D_dectector_inference_job = hub.submit_inference_job(
    model=bbox3D_dectector_target_model,
    device=device,
    inputs=bbox3D_dectector_input_data,
)
bbox3D_dectector_inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.
QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on 3D-Deep-BOX's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of 3D-Deep-BOX can be found here.
The license for the compiled assets for on-device deployment can be found here

References

Community

Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
For questions or feedback please reach out to us.