Skip to content
Switch branches/tags

YolactEdge: Real-time Instance Segmentation on the Edge

██╗   ██╗ ██████╗ ██╗      █████╗  ██████╗████████╗    ███████╗██████╗  ██████╗ ███████╗
╚██╗ ██╔╝██╔═══██╗██║     ██╔══██╗██╔════╝╚══██╔══╝    ██╔════╝██╔══██╗██╔════╝ ██╔════╝
 ╚████╔╝ ██║   ██║██║     ███████║██║        ██║       █████╗  ██║  ██║██║  ███╗█████╗  
  ╚██╔╝  ██║   ██║██║     ██╔══██║██║        ██║       ██╔══╝  ██║  ██║██║   ██║██╔══╝  
   ██║   ╚██████╔╝███████╗██║  ██║╚██████╗   ██║       ███████╗██████╔╝╚██████╔╝███████╗
   ╚═╝    ╚═════╝ ╚══════╝╚═╝  ╚═╝ ╚═════╝   ╚═╝       ╚══════╝╚═════╝  ╚═════╝ ╚══════╝

YolactEdge, the first competitive instance segmentation approach that runs on small edge devices at real-time speeds. Specifically, YolactEdge runs at up to 30.8 FPS on a Jetson AGX Xavier (and 172.7 FPS on an RTX 2080 Ti) with a ResNet-101 backbone on 550x550 resolution images. This is the code for our paper.

For a real-time demo and more samples, check out our demo video.




Model Zoo

We provide baseline YOLACT and YolactEdge models trained on COCO and YouTube VIS (our sub-training split, with COCO joint training).

To evalute the model, put the corresponding weights file in the ./weights directory and run one of the following commands.

YouTube VIS models:

Method Backbone  mAP AGX-Xavier FPS RTX 2080 Ti FPS weights
YOLACT R-50-FPN 44.7 8.5 59.8 download | mirror
(w/o TRT)
R-50-FPN 44.2 10.5 67.0 download | mirror
YolactEdge R-50-FPN 44.0 32.4 177.6 download | mirror
YOLACT R-101-FPN 47.3 5.9 42.6 download | mirror
(w/o TRT)
R-101-FPN 46.9 9.5 61.2 download | mirror
YolactEdge R-101-FPN 46.2 30.8 172.7 download | mirror

COCO models:

Method    Backbone     mAP Titan Xp FPS AGX-Xavier FPS RTX 2080 Ti FPS weights
YOLACT MobileNet-V2 22.1 - 15.0 35.7 download | mirror
YolactEdge MobileNet-V2 20.8 - 35.7 161.4 download | mirror
YOLACT R-50-FPN 28.2 42.5 9.1 45.0 download | mirror
YolactEdge R-50-FPN 27.0 - 30.7 140.3 download | mirror
YOLACT R-101-FPN 29.8 33.5 6.6 36.5 download | mirror
YolactEdge R-101-FPN 29.5 - 27.3 124.8 download | mirror



Optionally, you can use the official Dockerfile to set up full enivronment with one command.

Getting Started

Follow the installation instructions to set up required environment for running YolactEdge.

See instructions to evaluate and train with YolactEdge.

Colab Notebook

Try out our Colab Notebook with a live demo to learn about basic usage.

If you are interested in evaluating YolactEdge with TensorRT, we provide another Colab Notebook with TensorRT environment configuration on Colab.


Quantitative Results

# Convert each component of the trained model to TensorRT using the optimal settings and evaluate on the YouTube VIS validation set (our split).
python3 --trained_model=./weights/yolact_edge_vid_847_50000.pth

# Evaluate on the entire COCO validation set.
python3 --trained_model=./weights/yolact_edge_54_800000.pth

# Output a COCO JSON file for the COCO test-dev. The command will create './results/bbox_detections.json' and './results/mask_detections.json' for detection and instance segmentation respectively. These files can then be submitted to the website for evaluation.
python3 --trained_model=./weights/yolact_edge_54_800000.pth --dataset=coco2017_testdev_dataset --output_coco_json

Qualitative Results

# Display qualitative results on COCO. From here on I'll use a confidence threshold of 0.3.
python --trained_model=weights/yolact_edge_54_800000.pth --score_threshold=0.3 --top_k=100 --display


# Benchmark the trained model on the COCO validation set.
# Run just the raw model on the first 1k images of the validation set
python --trained_model=weights/yolact_edge_54_800000.pth --benchmark --max_images=1000


Handling inference error when using TensorRT

If you are using TensorRT conversion of YolactEdge and encountered issue in PostProcessing or NMS stage, this might be related to TensorRT engine issues. We implemented a experimental safe mode that will handle these cases carefully. Try this out with --use_tensorrt_safe_mode option in your command.

Inference using models trained with YOLACT

If you have a pre-trained model with YOLACT, and you want to take advantage of either TensorRT feature of YolactEdge, simply specify the --config=yolact_edge_config in command line options, and the code will automatically detect and convert the model weights to be compatible.

python3 --config=yolact_edge_config --trained_model=./weights/yolact_base_54_800000.pth

Inference without Calibration

If you want to run inference command without calibration, you can either run with FP16-only TensorRT optimization, or without TensorRT optimization with corresponding configs. Refer to data/ for examples of such configs.

# Evaluate YolactEdge with FP16-only TensorRT optimization with '--use_fp16_tensorrt' option (replace all INT8 optimization with FP16).
python3 --use_fp16_tensorrt --trained_model=./weights/yolact_edge_54_800000.pth

# Evaluate YolactEdge without TensorRT optimization with '--disable_tensorrt' option.
python3 --disable_tensorrt --trained_model=./weights/yolact_edge_54_800000.pth


# Display qualitative results on the specified image.
python --trained_model=weights/yolact_edge_54_800000.pth --score_threshold=0.3 --top_k=100 --image=my_image.png

# Process an image and save it to another file.
python --trained_model=weights/yolact_edge_54_800000.pth --score_threshold=0.3 --top_k=100 --image=input_image.png:output_image.png

# Process a whole folder of images.
python --trained_model=weights/yolact_edge_54_800000.pth --score_threshold=0.3 --top_k=100 --images=path/to/input/folder:path/to/output/folder


# Display a video in real-time. "--video_multiframe" will process that many frames at once for improved performance.
# If video_multiframe > 1, then the trt_batch_size should be increased to match it or surpass it. 
python --trained_model=weights/yolact_edge_54_800000.pth --score_threshold=0.3 --top_k=100 --video_multiframe=2 --trt_batch_size 2 --video=my_video.mp4

# Display a webcam feed in real-time. If you have multiple webcams pass the index of the webcam you want instead of 0.
python --trained_model=weights/yolact_edge_54_800000.pth --score_threshold=0.3 --top_k=100 --video_multiframe=2 --trt_batch_size 2 --video=0

# Process a video and save it to another file. This is unoptimized.
python --trained_model=weights/yolact_edge_54_800000.pth --score_threshold=0.3 --top_k=100 --video=input_video.mp4:output_video.mp4

Use the help option to see a description of all available command line arguments:

python --help


Make sure to download the entire dataset using the commands above.

  • To train, grab an imagenet-pretrained model and put it in ./weights.
    • For Resnet101, download resnet101_reducedfc.pth from here.
    • For Resnet50, download resnet50-19c8e357.pth from here.
    • For MobileNetV2, download mobilenet_v2-b0353104.pth from here.
  • Run one of the training commands below.
    • Note that you can press ctrl+c while training and it will save an *_interrupt.pth file at the current iteration.
    • All weights are saved in the ./weights directory by default with the file name <config>_<epoch>_<iter>.pth.
# Trains using the base edge config with a batch size of 8 (the default).
python --config=yolact_edge_config

# Resume training yolact_edge with a specific weight file and start from the iteration specified in the weight file's name.
python --config=yolact_edge_config --resume=weights/yolact_edge_10_32100.pth --start_iter=-1

# Use the help option to see a description of all available command line arguments
python --help

Training on video dataset

# Pre-train the image based model
python --config=yolact_edge_youtubevis_config

# Train the flow (warping) module
python --config=yolact_edge_vid_trainflow_config --resume=./weights/yolact_edge_youtubevis_847_50000.pth

# Fine tune the network jointly
python --config=yolact_edge_vid_config --resume=./weights/yolact_edge_vid_trainflow_144_100000.pth

Custom Datasets

You can also train on your own dataset by following these steps:

  • Depending on the type of your dataset, create a COCO-style (image) or YTVIS-style (video) Object Detection JSON annotation file for your dataset. The specification for this can be found here for COCO and YTVIS respectively. Note that we don't use some fields, so the following may be omitted:
    • info
    • liscense
    • Under image: license, flickr_url, coco_url, date_captured
    • categories (we use our own format for categories, see below)
  • Create a definition for your dataset under dataset_base in data/ (see the comments in dataset_base for an explanation of each field):
my_custom_dataset = dataset_base.copy({
    'name': 'My Dataset',

    'train_images': 'path_to_training_images',
    'train_info':   'path_to_training_annotation',

    'valid_images': 'path_to_validation_images',
    'valid_info':   'path_to_validation_annotation',

    'has_gt': True,
    'class_names': ('my_class_id_1', 'my_class_id_2', 'my_class_id_3', ...),

    # below is only needed for YTVIS-style video dataset.

    # whether samples all frames or key frames only.
    'use_all_frames': False,

    # the following four lines define the frame sampling strategy for the given dataset.
    'frame_offset_lb': 1,
    'frame_offset_ub': 4,
    'frame_offset_multiplier': 1,
    'all_frame_direction': 'allway',

    # 1 of K frames is annotated
    'images_per_video': 5,

    # declares a video dataset
    'is_video': True
  • Note that: class IDs in the annotation file should start at 1 and increase sequentially on the order of class_names. If this isn't the case for your annotation file (like in COCO), see the field label_map in dataset_base.
  • Finally, in yolact_edge_config in the same file, change the value for 'dataset' to 'my_custom_dataset' or whatever you named the config object above and 'num_classes' to number of classes in your dataset+1. Then you can use any of the training commands in the previous section.


If you use this code base in your work, please consider citing:

  author    = {Haotian Liu and Rafael A. Rivera Soto and Fanyi Xiao and Yong Jae Lee},
  title     = {YolactEdge: Real-time Instance Segmentation on the Edge},
  booktitle = {ICRA},
  year      = {2021},
  author    = {Daniel Bolya and Chong Zhou and Fanyi Xiao and Yong Jae Lee},
  title     = {YOLACT: {Real-time} Instance Segmentation},
  booktitle = {ICCV},
  year      = {2019},


For questions about our paper or code, please contact Haotian Liu or Rafael A. Rivera-Soto.


The first competitive instance segmentation approach that runs on small edge devices at real-time speeds.





No releases published


No packages published