README.md

---
library_name: pytorch
license: other
tags:
- bu_auto
- android
pipeline_tag: image-segmentation

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deeplab_xception/web-assets/model_demo.png)

# DeepLabXception: Optimized for Qualcomm Devices

DeepLabXception is a semantic segmentation model supporting multiple backbones like ResNet-101 and Xception, with flexible dataset compatibility including COCO, VOC, and Cityscapes.

This is based on the implementation of DeepLabXception found [here](https://github.com/LikeLy-Journey/SegmenTron).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/deeplab_xception) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).

Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) to run these models on a hosted Qualcomm® device.

## Getting Started
There are two ways to deploy this model on your device:

### Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.42, ONNX Runtime 1.24.3 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deeplab_xception/releases/v0.52.0/deeplab_xception-onnx-float.zip)
| ONNX | w8a8 | Universal | QAIRT 2.42, ONNX Runtime 1.24.3 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deeplab_xception/releases/v0.52.0/deeplab_xception-onnx-w8a8.zip)
| QNN_DLC | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deeplab_xception/releases/v0.52.0/deeplab_xception-qnn_dlc-float.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deeplab_xception/releases/v0.52.0/deeplab_xception-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deeplab_xception/releases/v0.52.0/deeplab_xception-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.45 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/deeplab_xception/releases/v0.52.0/deeplab_xception-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[DeepLabXception on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/deeplab_xception)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/deeplab_xception) Python library to compile and export the model with your own:
- Custom weights (e.g., fine-tuned checkpoints)
- Custom input shapes
- Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for [DeepLabXception on GitHub](https://github.com/qualcomm/ai-hub-models/blob/main/src/qai_hub_models/models/deeplab_xception) for usage instructions.

## Model Details

**Model Type:** Model_use_case.semantic_segmentation

**Model Stats:**
- Model checkpoint: COCO_WITH_VOC_LABELS_V1
- Input resolution: 480x520
- Number of output classes: 21
- Number of parameters: 41.26M
- Model size (float): 158 MB

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| DeepLabXception | ONNX | float | Snapdragon® 8 Elite Gen 5 Mobile | 8.886 ms | 3 - 255 MB | NPU
| DeepLabXception | ONNX | float | Snapdragon® X2 Elite | 10.437 ms | 86 - 86 MB | NPU
| DeepLabXception | ONNX | float | Snapdragon® X Elite | 22.96 ms | 85 - 85 MB | NPU
| DeepLabXception | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 17.017 ms | 0 - 366 MB | NPU
| DeepLabXception | ONNX | float | Qualcomm® QCS8550 (Proxy) | 22.768 ms | 3 - 6 MB | NPU
| DeepLabXception | ONNX | float | Qualcomm® QCS9075 | 32.928 ms | 3 - 8 MB | NPU
| DeepLabXception | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 13.216 ms | 1 - 242 MB | NPU
| DeepLabXception | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 3.056 ms | 0 - 293 MB | NPU
| DeepLabXception | ONNX | w8a8 | Snapdragon® X2 Elite | 3.344 ms | 44 - 44 MB | NPU
| DeepLabXception | ONNX | w8a8 | Snapdragon® X Elite | 8.058 ms | 44 - 44 MB | NPU
| DeepLabXception | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 5.278 ms | 0 - 391 MB | NPU
| DeepLabXception | ONNX | w8a8 | Qualcomm® QCS6490 | 311.631 ms | 90 - 116 MB | CPU
| DeepLabXception | ONNX | w8a8 | Qualcomm® QCS8550 (Proxy) | 7.435 ms | 0 - 99 MB | NPU
| DeepLabXception | ONNX | w8a8 | Qualcomm® QCS9075 | 7.603 ms | 0 - 4 MB | NPU
| DeepLabXception | ONNX | w8a8 | Qualcomm® QCM6690 | 243.471 ms | 117 - 129 MB | CPU
| DeepLabXception | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 4.212 ms | 0 - 280 MB | NPU
| DeepLabXception | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 229.688 ms | 90 - 103 MB | CPU
| DeepLabXception | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 Mobile | 9.538 ms | 3 - 256 MB | NPU
| DeepLabXception | QNN_DLC | float | Snapdragon® X2 Elite | 11.531 ms | 3 - 3 MB | NPU
| DeepLabXception | QNN_DLC | float | Snapdragon® X Elite | 22.866 ms | 3 - 3 MB | NPU
| DeepLabXception | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 16.073 ms | 0 - 355 MB | NPU
| DeepLabXception | QNN_DLC | float | Qualcomm® QCS8275 (Proxy) | 112.912 ms | 1 - 245 MB | NPU
| DeepLabXception | QNN_DLC | float | Qualcomm® QCS8550 (Proxy) | 22.124 ms | 3 - 4 MB | NPU
| DeepLabXception | QNN_DLC | float | Qualcomm® SA8775P | 30.612 ms | 1 - 245 MB | NPU
| DeepLabXception | QNN_DLC | float | Qualcomm® QCS9075 | 35.912 ms | 3 - 8 MB | NPU
| DeepLabXception | QNN_DLC | float | Qualcomm® QCS8450 (Proxy) | 45.848 ms | 0 - 352 MB | NPU
| DeepLabXception | QNN_DLC | float | Qualcomm® SA7255P | 112.912 ms | 1 - 245 MB | NPU
| DeepLabXception | QNN_DLC | float | Qualcomm® SA8295P | 39.128 ms | 0 - 241 MB | NPU
| DeepLabXception | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 12.19 ms | 0 - 251 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 3.242 ms | 1 - 280 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 3.873 ms | 1 - 1 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Snapdragon® X Elite | 8.525 ms | 1 - 1 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 5.55 ms | 1 - 369 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® QCS6490 | 33.692 ms | 1 - 3 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® QCS8275 (Proxy) | 20.37 ms | 1 - 282 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® QCS8550 (Proxy) | 7.915 ms | 1 - 3 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® SA8775P | 8.024 ms | 1 - 285 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® QCS9075 | 8.62 ms | 3 - 5 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® QCM6690 | 94.033 ms | 1 - 304 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® QCS8450 (Proxy) | 13.928 ms | 1 - 367 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® SA7255P | 20.37 ms | 1 - 282 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Qualcomm® SA8295P | 12.732 ms | 0 - 278 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 4.347 ms | 1 - 290 MB | NPU
| DeepLabXception | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 12.458 ms | 1 - 278 MB | NPU
| DeepLabXception | TFLITE | float | Snapdragon® 8 Elite Gen 5 Mobile | 9.188 ms | 0 - 357 MB | NPU
| DeepLabXception | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 16.193 ms | 0 - 455 MB | NPU
| DeepLabXception | TFLITE | float | Qualcomm® QCS8275 (Proxy) | 113.123 ms | 0 - 343 MB | NPU
| DeepLabXception | TFLITE | float | Qualcomm® QCS8550 (Proxy) | 21.226 ms | 0 - 3 MB | NPU
| DeepLabXception | TFLITE | float | Qualcomm® SA8775P | 30.643 ms | 0 - 343 MB | NPU
| DeepLabXception | TFLITE | float | Qualcomm® QCS9075 | 35.513 ms | 0 - 93 MB | NPU
| DeepLabXception | TFLITE | float | Qualcomm® QCS8450 (Proxy) | 46.024 ms | 0 - 448 MB | NPU
| DeepLabXception | TFLITE | float | Qualcomm® SA7255P | 113.123 ms | 0 - 343 MB | NPU
| DeepLabXception | TFLITE | float | Qualcomm® SA8295P | 38.721 ms | 0 - 337 MB | NPU
| DeepLabXception | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 12.249 ms | 0 - 346 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 Mobile | 2.88 ms | 0 - 284 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 4.978 ms | 0 - 369 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® QCS6490 | 31.905 ms | 0 - 58 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® QCS8275 (Proxy) | 19.029 ms | 0 - 284 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® QCS8550 (Proxy) | 6.979 ms | 0 - 2 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® SA8775P | 7.258 ms | 0 - 284 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® QCS9075 | 7.479 ms | 0 - 47 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® QCM6690 | 122.211 ms | 0 - 294 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® QCS8450 (Proxy) | 11.946 ms | 0 - 358 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® SA7255P | 19.029 ms | 0 - 284 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Qualcomm® SA8295P | 11.629 ms | 0 - 275 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 3.929 ms | 0 - 274 MB | NPU
| DeepLabXception | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 11.166 ms | 0 - 266 MB | NPU

## License
* The license for the original implementation of DeepLabXception can be found
  [here](https://github.com/pytorch/vision/blob/main/LICENSE).

## References
* [Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation](https://arxiv.org/abs/1802.02611)
* [Source Model Implementation](https://github.com/LikeLy-Journey/SegmenTron)

## Community
* Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
* For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com).