EdgeCrafter: compact ViTs for detection, segmentation and pose on edge hardware

What EdgeCrafter is for, and who should care

Dense prediction tasks such as object detection, instance segmentation and pose estimation normally run on server-class GPUs. EdgeCrafter targets the opposite end: compact vision transformers that fit on edge accelerators while keeping accuracy close to larger models. The repository is the official implementation of a TMLR 2026 paper, and the model zoo covers three task families under one naming scheme: ECDet for detection, ECSeg for instance segmentation and ECPose for pose estimation, each in S, M, L and X sizes.

The intended user is an engineer who has a fixed latency budget on a T4-class or smaller device and cannot afford a full-size backbone. The README states that latency figures were measured on an NVIDIA T4 GPU with batch size 1 under FP16 using TensorRT v10.6, which tells you the numbers are deployment-oriented rather than training-oriented. If your workload runs on a data centre GPU with no latency constraint, the compactness buys you nothing and a larger model is the simpler choice.

How task-specialized distillation shapes the architecture

The paper title names the mechanism: task-specialized distillation. Rather than training one compact model to imitate a general teacher, EdgeCrafter trains each compact variant against the signal that matters for its task. The repository structure reflects this. There is no single training entry point. ecdetseg holds the detection and instance segmentation code and configs, while ecpose holds the pose estimation code. The two directories are separate reproduction paths, and the README links to each as its own set of instructions.

The backbone lineage is visible in the topics list, which includes dinov3. The model zoo also documents a second axis of variation: COCO-only training versus additional Objects365 pretraining, marked in the tables as `--` and `O365`. For ECSeg and ECPose, the README notes that the O365 results come from transferring the detection model pretrained on Objects365 rather than from a separate segmentation or pose pretraining run. That is a real design decision with a cost: the segmentation and pose variants inherit whatever the detection pretraining learned, which is efficient but couples the three task families to one pretraining corpus.

Installing EdgeCrafter and running a first detection config

The repository does not document a pip package. Installation is a clone plus the pinned requirements file at the repository root. The requirements pin several versions tightly, including numpy 2.0.1, scipy 1.16.0 and onnx 1.19.0, so a fresh virtual environment is the safer starting point than an existing one.

git clone https://github.com/Intellindust-AI-Lab/EdgeCrafter.git
cd EdgeCrafter
pip install -r requirements.txt

After that, the README points to the ecdetseg directory for detection and instance segmentation instructions. The configs live under ecdetseg/configs/ecdet/ for detection and ecdetseg/configs/ecseg/ for segmentation, with one YAML per size. The smallest detection config is ecdetseg/configs/ecdet/ecdet_s.yml.

cd ecdetseg
# follow the instructions in this directory for training and evaluation

The README does not print the exact training command in the top-level file, so read the instructions inside ecdetseg before assuming a flag. Checkpoints are distributed as release assets rather than through a package index, and the README also points to a Hugging Face organization for models, with hf_models.ipynb at the repository root as a notebook path to them. If you only want to try inference, the Hugging Face notebook is the shortest route; if you want to reproduce the reported numbers, you need the config and the matching checkpoint from the same release tag.

Licence and the cost of staying current

The repository metadata reports the licence as NOASSERTION, and the file list contains LICENSE.md. The README badge links to a file called LICENSE and labels it EdgeCrafter License. That mismatch between the badge target, the actual filename and the metadata classification is worth resolving before you ship anything. A custom licence is not automatically restrictive, but it is not Apache 2.0 or MIT either, and the terms are not summarised in the README. Read LICENSE.md directly and decide whether your distribution model fits. Nothing here is legal advice.

On maintenance, the last push was on 2026-08-24. The updates list shows a steady cadence through 2026, including Objects365 checkpoints on 2026-08-14 and an Intel Geti integration on 2026-08-13. There are no retrieved releases in the metadata, yet the README links checkpoints under release tags such as edgecrafterv1 and edgecrafterv1_o365, so the release assets are the practical upgrade unit. Upgrading means re-downloading checkpoints and re-checking the config that produced them, because a config in ecdetseg/configs/ecdet/ and a checkpoint from a different tag are not guaranteed to line up. The pinned requirements add a second cost: a PyTorch upgrade beyond the torch >= 2.6.0 floor may pull numpy or scipy past their pins, and the README does not document a tested combination beyond what requirements.txt states.

Where EdgeCrafter is the wrong tool

The model zoo is entirely COCO and Objects365 oriented. If your domain is medical imaging, aerial imagery or industrial defect detection, the pretrained checkpoints give you a starting point but not a solution, and the repository's reproduction instructions assume the standard benchmarks. Fine-tuning on your own data is possible, and the Intel Geti integration announced on 2026-08-13 is described as a no-code path for fine-tuning ECDet-S/M/L/X, but that route runs through a separate product rather than through this repository.

The second limitation is documentation granularity. The top-level README is a model zoo and a set of links. The actual commands live in ecdetseg and ecpose, which the README does not inline. If you want a single file that tells you how to train and evaluate end to end, this is not it. The third is the licence, covered above. A team that needs a permissively licensed detector for a commercial product has to read LICENSE.md before writing any integration code, and may find the terms unsuitable.

How EdgeCrafter differs from DEIMv2 and RT-DETR-style detectors

The most direct comparison is DEIMv2, from the same lab. The README describes DEIMv2 as the previous version and notes that it was used by two winning teams at the CVPR 2026 Maritime Computer Vision Workshop. DEIMv2 is a detection line. EdgeCrafter extends the same lab's direction to three dense prediction tasks under one compact ViT family, with distillation as the stated training mechanism and DINOv3 in the backbone lineage. If you only need detection, DEIMv2 is the narrower and more established option; if you need segmentation and pose from a consistent family, EdgeCrafter is the one that covers all three.

Against RT-DETR-style real-time detectors, the difference is the training signal rather than the task head. Those models are typically trained on labelled data alone or with a general-purpose distillation setup. EdgeCrafter's stated approach is task-specialized distillation, meaning the teacher signal is chosen per task. Whether that yields a better accuracy-per-latency trade on your hardware is something the README answers only for a T4 under FP16 with TensorRT v10.6. On a different accelerator, the latency column does not transfer, and you should measure rather than extrapolate.

Editorial conclusion

Adopt EdgeCrafter if you need a single compact ViT family covering detection, instance segmentation and pose estimation, and you are comfortable cloning two subdirectories for instructions. Do not adopt it if you need a permissive OSI licence, because the repository ships a custom EdgeCrafter License rather than Apache or MIT. Before committing, verify the licence terms against your product, confirm the checkpoint you want exists in the release assets, and check that the environment matches the pinned versions in requirements.txt rather than a newer PyTorch.

Frequently asked questions

Official sources

1. Intellindust-AI-Lab/EdgeCrafter on GitHub
2. Issues
3. Project website
4. README