[TMLR 26] EdgeCrafter: Compact ViTs for Edge Dense Prediction via Task-Specialized Distillation

English | 简体中文

Longfei Liu ^*‡  Yongjie Hou ^*  Yang Li ^*  Qirui Wang ^*  Youyang Sha 
Yongjun Yu  Yinzhi Wang  Peizhe Ru  Xuanlong Yu^†  Xi Shen ^†

* Equal Contribution    ‡ Project Lead    † Corresponding Author

Intellindust AI Lab

---

🚀 Updates

• [2026-08-14] We released the Objects365 pretrained checkpoints. Enjoy it!
• [2026-08-13] EdgeCrafter has been integrated into Intel® Geti™: fine-tune ECDet-S/M/L/X on your own data with no code, train on Intel dGPU / iGPU / CPU, export to OpenVINO™ IR and optimized for INT8 for edge deployment.
• [2026-08-11] EdgeCrafter has been accepted for publication in TMLR 2026 🎉
• [2026-07-23] EdgeCrafter has been integrated into LightlyTrain
• [2026-07-22] EdgeCrafter has been integrated into LibreYOLO.
• [2026-04-20] Our previous version DEIMv2 was used by two winning teams at the CVPR 2026 Maritime Computer Vision Workshop, taking home 2nd place in Thermal Object Detection Challenge and 3rd place in Vision-to-Chart Data Association Challenge.
• [2026-03-21] We have uploaded our models on 🤗 Hugging Face.
• [2026-03-19] Initial release of EdgeCrafter.

---

🤗 Hugging Face

We have uploaded our models on 🤗 Hugging Face! You can also access these models via hf_models.ipynb. Have a try!

---

📍 Reproducing the Results

• Detection & Instance Segmentation: Instructions
• Pose Estimation: Instructions

---

🏆 Model Zoo

COCO2017 Validation Results

Note: Latency is measured on an NVIDIA T4 GPU with batch size 1 under FP16 precision using TensorRT (v10.6).

-- denotes COCO-only training, and O365 denotes additional Objects365 pretraining. For ECSeg and ECPose, the O365 results are obtained by directly transferring the detection model pretrained on Objects365.

Object Detection

Model	Extra Sup.	Size	AP50:95	#Params	GFLOPs	Latency (ms)	Config	Log	Checkpoint
ECDet-S	--	640	51.7	10	26	5.41	config	log	model
ECDet-S	O365	640	53.6	10	26	5.41	-	-	model
ECDet-M	--	640	54.3	19	53	7.98	config	log	model
ECDet-M	O365	640	56.7	19	53	7.98	-	-	model
ECDet-L	--	640	57.0	33	101	10.49	config	log	model
ECDet-L	O365	640	59.0	33	101	10.49	-	-	model
ECDet-X	--	640	57.9	49	151	12.70	config	log	model
ECDet-X	O365	640	59.9	49	151	12.70	-	-	model

Objects365 Pretrained Checkpoints

Model	Checkpoint
ECDet-S	model
ECDet-M	model
ECDet-L	model
ECDet-X	model

Instance Segmentation

Model	Extra Sup.	Size	AP50:95	#Params	GFLOPs	Latency (ms)	Config	Log	Checkpoint
ECSeg-S	--	640	43.0	10	33	6.96	config	log	model
ECSeg-S	O365	640	43.9	10	33	6.96	-	-	model
ECSeg-M	--	640	45.2	20	64	9.85	config	log	model
ECSeg-M	O365	640	46.9	20	64	9.85	-	-	model
ECSeg-L	--	640	47.1	34	111	12.56	config	log	model
ECSeg-L	O365	640	48.8	34	111	12.56	-	-	model
ECSeg-X	--	640	48.4	50	168	14.96	config	log	model
ECSeg-X	O365	640	49.8	50	168	14.96	-	-	model

Pose Estimation

Model	Extra Sup.	Size	AP50:95	#Params	GFLOPs	Latency (ms)	Config	Log	Checkpoint
ECPose-S	--	640	68.9	10	30	5.54	config	log	model
ECPose-S	O365	640	69.7	10	30	5.54	-	-	model
ECPose-M	--	640	72.4	20	63	9.25	config	log	model
ECPose-M	O365	640	73.1	20	63	9.25	-	-	model
ECPose-L	--	640	73.5	34	112	11.83	config	log	model
ECPose-L	O365	640	74.5	34	112	11.83	-	-	model
ECPose-X	--	640	74.8	51	172	14.31	config	log	model
ECPose-X	O365	640	75.9	51	172	14.31	-	-	model

---

📦 Installation

# Create conda environment
conda create -n ec python=3.11 -y
conda activate ec

# Install dependencies
pip install -r requirements.txt

⚡ Quick Start (Inference)

The easiest way to test EdgeCrafter is to run inference on a sample image using a pre-trained model.

# 1. Download a pre-trained model (e.g., ECDet-L)
cd ecdetseg
wget https://github.com/capsule2077/edgecrafter/releases/download/edgecrafterv1/ecdet_l.pth
# 2. Run PyTorch inference
# Make sure to replace `path/to/your/image.jpg` with an actual image path
python tools/inference/torch_inf.py -c configs/ecdet/ecdet_l.yml -r ecdet_l.pth -i path/to/your/image.jpg

📄 License

This project is released under the Apache 2.0 License.

---

🙏 Acknowledgements

We thank the authors of the following open-source projects that made this work possible: RT-DETR, D-FINE, DEIM, lightly-train, DETRPose, RF-DETR, DINOv3

---

📚 Citation

If you find this project useful in your research, please consider citing:

@article{liu2026edgecrafter,
  title={EdgeCrafter: Compact ViTs for Edge Dense Prediction via Task-Specialized Distillation},
  author={Liu, Longfei and Hou, Yongjie and Li, Yang and Wang, Qirui and Sha, Youyang and Yu, Yongjun and Wang, Yinzhi and Ru, Peizhe and Yu, Xuanlong and Shen, Xi},
  journal={TMLR},
  year={2026}
}