CutLER¶

Development line: project:cutler · thread cutler-development
Last event: 2024-03-11 · 2 dated since 2023-03-01 · Researched: 2026-09-04 · confidence: medium

What it is¶

CutLER is a pseudo-mask-to-detector pipeline for unsupervised object detection and instance segmentation.

• MaskCut generates multiple-instance pseudo-masks.
• CutLER trains image detectors and segmentors from those masks.
• VideoCutLER turns pairs of images into synthetic video trajectories, then trains video instance segmentation without human labels, natural video, or optical flow.

The setup targets Linux/macOS, Python 3.8+, PyTorch 1.8+, Detectron2, and OpenCV 4.6+; the repository has no tagged releases.

We can use it as research code or a reproducible baseline. Smoke-test the legacy dependency stack before committing to training.

Development line¶

• 2023-03-01 — CutLER project resources became available. MaskCut produces pseudo-masks, then CutLER learns object detection and instance segmentation from ImageNet-1K without human annotations.
• 2024-03-11 — VideoCutLER documentation extended the CutLER project line. MaskCut → ImageCut2Video → video-model training, extending the project to tracking multiple instances across frames.

What changed¶

2023-03-01 — CutLER established the image workflow: MaskCut produces pseudo-masks, then CutLER learns object detection and instance segmentation from ImageNet-1K without human annotations.

2023-08-27–29 — The official Git history places the VideoCutLER code addition before the later documentation event, in commits labelled “Add videocutler.”

2024-03-02 — Official history records two further VideoCutLER commits, but their terse messages do not establish a specific functional delta.

2024-03-11 — VideoCutLER is documented as a separate video workflow: MaskCut → ImageCut2Video → video-model training, extending the project to tracking multiple instances across frames.

2025-06-03–04 — Upstream updated VideoCutLER checkpoint links and its README. No tagged release is available.

How to use this¶

From 2023-03-01, practitioners could use the CutLER repository, MaskCut space, and Colab resources to evaluate the project. From 2024-03-11, consult the VideoCutLER documentation when the use case involves video.

1. Install the documented local stack: Linux or macOS, compatible PyTorch/torchvision and Detectron2. Clone CutLER recursively and install its requirements. — https://github.com/facebookresearch/CutLER/blob/main/INSTALL.md
2. For still images, start with the local MaskCut or pretrained CutLER demo. Select the matching model-zoo config and checkpoint, then save outputs locally. — https://github.com/facebookresearch/CutLER
3. For video inference, download the official VideoCutLER checkpoint. Run videocutler/demo_video/demo.py on sequential image frames with the supplied Mask2Former config, an output directory, and model weights. — https://github.com/facebookresearch/CutLER/blob/main/videocutler/README.md
4. For retraining, prepare ImageNet-1K, then generate or download MaskCut pseudo-masks. Place the CutLER pretrain in videocutler/pretrain, then run train_net_video.py. — https://github.com/facebookresearch/CutLER/blob/main/videocutler/README.md

Best practices¶

• Keep PyTorch, torchvision, and Detectron2 versions mutually compatible. The upstream example uses PyTorch 1.8.1, so treat a modern environment as a separate validation target. — https://github.com/facebookresearch/CutLER/blob/main/INSTALL.md
• Use the local workflow instead of the MaskCut web demo: the official Hugging Face Space currently reports a build error. — https://huggingface.co/spaces/facebook/MaskCut
• For VideoCutLER, begin with the documented 0.8 confidence threshold. Lower it only when higher recall is the priority, and save frames or masks for inspection. — https://github.com/facebookresearch/CutLER/blob/main/videocutler/README.md
• For large MaskCut annotation runs, shard folders with --num-folder-per-job and --job-index. Merge JSON files using matching fixed_size, tau, and N settings. — https://github.com/facebookresearch/CutLER
• Pin the exact repository commit and checkpoint used for a run, because upstream provides no tagged releases. — https://github.com/facebookresearch/CutLER/tags

Superseded by this¶

• 2023-03-01: the MaskCut web-demo route is obsolete for current use. Its official Hugging Face Space now reports a build error, so use local code instead.
• 2023-03-01: image-only CutLER guidance is incomplete for multi-instance video work. Use VideoCutLER's dedicated inference and training workflow while retaining CutLER for still images.
• 2024-03-11: stale VideoCutLER checkpoint documentation should yield to the upstream README revision recorded on 2025-06-03–04. Checkpoint availability was not independently downloaded.

Still unknown¶

• We did not install or run the repository; compatibility with current CUDA, PyTorch, and Detectron2 versions remains unverified.
• The original Berkeley project page and Colab notebook from 2023 could not be independently retrieved or run in this session, so neither supports a current-use claim.
• The official history identifies the March 2024 VideoCutLER update but not its exact behavioral change.

Sources¶

Agent brief¶

• Subject: project:cutler, thread cutler-development, 2 dated events 2023-03-01 → 2024-03-11.
• Practical note: From 2023-03-01, practitioners could use the CutLER repository, MaskCut space, and Colab resources to evaluate the project; from 2024-03-11, they should also consult the VideoCutLER documentation when their use case involves video.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.