NanoSAM

👍 Usage - ⏱️ Performance - 🛠️ Setup - 🤸 Examples - 🏋️ Training
- 🧐 Evaluation - 👏 Acknowledgment - 🔗 See also

NanoSAM is a Segment Anything (SAM) model variant that is capable of running in 🔥 real-time 🔥 on NVIDIA Jetson Orin Platforms with NVIDIA TensorRT.

NanoSAM is trained by distilling the MobileSAM image encoder on unlabeled images. For an introduction to knowledge distillation, we recommend checking out this tutorial.

👍 Usage

Using NanoSAM from Python looks like this

from nanosam.utils.predictor import Predictor

predictor = Predictor(
    image_encoder="data/resnet18_image_encoder.engine",
    mask_decoder="data/mobile_sam_mask_decoder.engine"
)

image = PIL.Image.open("dog.jpg")

predictor.set_image(image)

mask, _, _ = predictor.predict(np.array([[x, y]]), np.array([1]))

Notes

The point labels may be

Point Label	Description
0	Background point
1	Foreground point
2	Bounding box top-left
3	Bounding box bottom-right

Follow the instructions below for how to build the engine files.

⏱️ Performance

NanoSAM runs real-time on Jetson Orin Nano.

Model †	⏱️ Jetson Orin Nano (ms)		⏱️ Jetson AGX Orin (ms)		🎯 Accuracy (mIoU) ‡
	Image Encoder	Full Pipeline	Image Encoder	Full Pipeline	All	Small	Medium	Large
MobileSAM	TBD	146	35	39	0.728	0.658	0.759	0.804
NanoSAM (ResNet18)	TBD	27	4.2	8.1	0.706	0.624	0.738	0.796

Notes

† The MobileSAM image encoder is optimized with FP32 precision because it produced erroneous results when built for FP16 precision with TensorRT. The NanoSAM image encoder is built with FP16 precision as we did not notice a significant accuracy degredation. Both pipelines use the same mask decoder which is built with FP32 precision. For all models, the accuracy reported uses the same model configuration used to measure latency.

‡ Accuracy is computed by prompting SAM with ground-truth object bounding box annotations from the COCO 2017 validation dataset. The IoU is then computed between the mask output of the SAM model for the object and the ground-truth COCO segmentation mask for the object. The mIoU is the average IoU over all objects in the COCO 2017 validation set matching the target object size (small, medium, large).

🛠️ Setup

NanoSAM is fairly easy to get started with.

1. Install the dependencies

   1. Install PyTorch

   2. Install torch2trt

   3. Install NVIDIA TensorRT

   4. (optional) Install TRTPose - For the pose example.

      git clone https://github.com/NVIDIA-AI-IOT/trt_pose
      cd trt_pose
      python3 setup.py develop --user

   5. (optional) Install the Transformers library - For the OWL ViT example.

      python3 -m pip install transformers

2. Install the NanoSAM Python package

   git clone https://github.com/NVIDIA-AI-IOT/nanosam
   cd nanosam
   python3 setup.py develop --user

3. Build the TensorRT engine for the mask decoder

   1. Export the MobileSAM mask decoder ONNX file (or download directly from here)

      python3 -m nanosam.tools.export_sam_mask_decoder_onnx \
          --model-type=vit_t \
          --checkpoint=assets/mobile_sam.pt \
          --output=data/mobile_sam_mask_decoder.onnx

   2. Build the TensorRT engine

      trtexec \
          --onnx=data/mobile_sam_mask_decoder.onnx \
          --saveEngine=data/mobile_sam_mask_decoder.engine \
          --minShapes=point_coords:1x1x2,point_labels:1x1 \
          --optShapes=point_coords:1x1x2,point_labels:1x1 \
          --maxShapes=point_coords:1x10x2,point_labels:1x10

      This assumes the mask decoder ONNX file is downloaded to data/mobile_sam_mask_decoder.onnx

      Notes

      This command builds the engine to support up to 10 keypoints. You can increase this limit as needed by specifying a different max shape.

4. Build the TensorRT engine for the NanoSAM image encoder

   1. Download the image encoder: resnet18_image_encoder.onnx

   2. Build the TensorRT engine

      trtexec \
          --onnx=data/resnet18_image_encoder.onnx \
          --saveEngine=data/resnet18_image_encoder.engine \
          --fp16

5. Run the basic usage example

   python3 examples/basic_usage.py \
       --image_encoder=data/resnet18_image_encoder.engine \
       --mask_decoder=data/mobile_sam_mask_decoder.engine
   

   This outputs a result to data/basic_usage_out.jpg

That's it! From there, you can read the example code for examples on how to use NanoSAM with Python. Or try running the more advanced examples below.

🤸 Examples

NanoSAM can be applied in many creative ways.

Example 1 - Segment with bounding box

This example uses a known image with a fixed bounding box to control NanoSAM segmentation.

To run the example, call

python3 examples/basic_usage.py \
    --image_encoder="data/resnet18_image_encoder.engine" \
    --mask_decoder="data/mobile_sam_mask_decoder.engine"

Example 2 - Segment with bounding box (using OWL-ViT detections)

This example demonstrates using OWL-ViT to detect objects using a text prompt(s), and then segmenting these objects using NanoSAM.

To run the example, call

python3 examples/segment_from_owl.py \
    --prompt="A tree" \
    --image_encoder="data/resnet18_image_encoder.engine" \
    --mask_decoder="data/mobile_sam_mask_decoder.engine

Notes

- While OWL-ViT does not run real-time on Jetson Orin Nano (3sec/img), it is nice for experimentation as it allows you to detect a wide variety of objects. You could substitute any other real-time pre-trained object detector to take full advantage of NanoSAM's speed.

Example 3 - Segment with keypoints (offline using TRTPose detections)

This example demonstrates how to use human pose keypoints from TRTPose to control NanoSAM segmentation.

To run the example, call

python3 examples/segment_from_pose.py

This will save an output figure to data/segment_from_pose_out.png.

Example 4 - Segment with keypoints (online using TRTPose detections)

This example demonstrates how to use human pose to control segmentation on a live camera feed. This example requires an attached display and camera.

To run the example, call

python3 examples/demo_pose_tshirt.py

Example 5 - Segment and track (experimental)

This example demonstrates a rudimentary segmentation tracking with NanoSAM. This example requires an attached display and camera.

To run the example, call

python3 examples/demo_click_segment_track.py <image_encoder_engine> <mask_decoder_engine>

Once the example is running double click an object you want to track.

Notes

This tracking method is very simple and can get lost easily. It is intended to demonstrate creative ways you can use NanoSAM, but would likely be improved with more work.

🏋️ Training

You can train NanoSAM on a single GPU

1. Download and extract the COCO 2017 train images

   # mkdir -p data/coco  # uncomment if it doesn't exist
   mkdir -p data/coco
   cd data/coco
   wget http://images.cocodataset.org/zips/train2017.zip
   unzip train2017.zip
   cd ../..

2. Build the MobileSAM image encoder (used as teacher model)

   1. Export to ONNX

      python3 -m nanosam.tools.export_sam_image_encoder_onnx \
          --checkpoint="assets/mobile_sam.pt" \
          --output="data/mobile_sam_image_encoder_bs16.onnx" \
          --model_type=vit_t \
          --batch_size=16

   2. Build the TensorRT engine with batch size 16

      trtexec \
          --onnx=data/mobile_sam_image_encoder_bs16.onnx \
          --shapes=image:16x3x1024x1024 \
          --saveEngine=data/mobile_sam_image_encoder_bs16.engine

3. Train the NanoSAM image encoder by distilling MobileSAM

   python3 -m nanosam.tools.train \
       --images=data/coco/train2017 \
       --output_dir=data/models/resnet18 \
       --model_name=resnet18 \
       --teacher_image_encoder_engine=data/mobile_sam_image_encoder_bs16.engine \
       --batch_size=16

   Notes

   Once training, visualizations of progress and checkpoints will be saved to the specified output directory. You can stop training and resume from the last saved checkpoint if needed.

   For a list of arguments, you can type

   python3 -m nanosam.tools.train --help

4. Export the trained NanoSAM image encoder to ONNX

   python3 -m nanosam.tools.export_image_encoder_onnx \
       --model_name=resnet18 \
       --checkpoint="data/models/resnet18/checkpoint.pth" \
       --output="data/resnet18_image_encoder.onnx"

You can then build the TensorRT engine as detailed in the getting started section.

🧐 Evaluation

You can reproduce the accuracy results above by evaluating against COCO ground truth masks

1. Download and extract the COCO 2017 validation set.

   # mkdir -p data/coco  # uncomment if it doesn't exist
   cd data/coco
   wget http://images.cocodataset.org/zips/val2017.zip
   wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
   unzip val2017.zip
   unzip annotations_trainval2017.zip
   cd ../..

2. Compute the IoU of NanoSAM mask predictions against the ground truth COCO mask annotation.

   python3 -m nanosam.tools.eval_coco \
       --coco_root=data/coco/val2017 \
       --coco_ann=data/coco/annotations/instances_val2017.json \
       --image_encoder=data/resnet18_image_encoder.engine \
       --mask_decoder=data/mobile_sam_mask_decoder.engine \
       --output=data/resnet18_coco_results.json

   This uses the COCO ground-truth bounding boxes as inputs to NanoSAM

3. Compute the average IoU over a selected category or size

   python3 -m nanosam.tools.compute_eval_coco_metrics \
       data/efficientvit_b0_coco_results.json \
       --size="all"

   Notes

   For all options type ``python3 -m nanosam.tools.compute_eval_coco_metrics --help``.

   To compute the mIoU for a specific category id.

   python3 -m nanosam.tools.compute_eval_coco_metrics \
       data/resnet18_coco_results.json \
       --category_id=1

👏 Acknowledgement

This project is enabled by the great projects below.

• SAM - The original Segment Anything model.
• MobileSAM - The distilled Tiny ViT Segment Anything model.

🔗 See also

• Jetson Introduction to Knowledge Distillation Tutorial - For an introduction to knowledge distillation as a model optimization technique.
• Jetson Generative AI Playground - For instructions and tips for using a variety of LLMs and transformers on Jetson.
• Jetson Containers - For a variety of easily deployable and modular Jetson Containers