Overview

Top-down instance segmentation normally follows a one-way detect-then-segment pipeline. Turbo-Inference closes the loop: coarse instance masks feed pixel-level localization and mask-quality information back to the detector, and the refined boxes are then used to predict better masks.

The method is training-free, reuses the pretrained mask head, and can be attached to existing top-down instance segmentation models without changing their training procedure.

The feedback loop contains three operations:

1. Box refinement maps each RoI mask back to image coordinates and derives a tighter box from its foreground support.
2. Maskness rescoring estimates mask quality from the foreground probability distribution and fuses it with the classification score.
3. Turbo segmentation extracts RoI features from the refined boxes and reuses the original mask head to predict more accurate masks.

Detection → coarse masks → refined boxes and scores → refined masks
                         ↖___________________________|

Highlights

• No retraining: all refinement happens during inference.
• Plug-and-play: the shared MaskBoxRefiner module is configurable from the RoI head.
• Joint improvement: both bounding-box AP and mask AP benefit from segmentation feedback.
• Broad applicability: experimental paths are provided for Mask R-CNN, Cascade Mask R-CNN/HTC, QueryInst/Sparse R-CNN, and RTMDet-Ins.
• Reproducible: the released Mask R-CNN implementation has been freshly evaluated on all 5,000 COCO 2017 validation images.
• Ready-to-use weights: the Checkpoint Zoo provides 18 pretrained models with original release links and matching configs.

Qualitative Results

Turbo-Inference produces tighter detection boxes, suppresses low-quality duplicate predictions, and improves instance masks.

Results

Main results on COCO

Turbo-Inference consistently improves representative two-stage, cascade, query-based, and one-stage instance segmentation frameworks. The following compact table reports the main results from the paper; FPS was measured on one RTX 2080 Ti with batch size 2.

Method	Backbone	bbox AP	bbox AP w/ Turbo	segm AP	segm AP w/ Turbo	FPS → Turbo FPS
Mask R-CNN	R50-FPN	39.2	40.3 (+1.1)	35.4	36.7 (+1.3)	15.7 → 12.0
HTC	R50-FPN	43.3	43.7 (+0.4)	38.3	39.2 (+0.9)	5.5 → 4.5
RTMDet-m	CSPX-PAFPN	48.8	49.3 (+0.5)	42.1	42.4 (+0.3)	3.7 → 2.7
QueryInst	R50-FPN	42.0	42.8 (+0.8)	37.5	38.7 (+1.2)	7.5 → 6.0

Results with additional backbones and model scales

Method	Backbone	bbox AP	bbox AP w/ Turbo	segm AP	segm AP w/ Turbo	FPS → Turbo FPS
Mask R-CNN	R101-FPN	40.8	41.8	36.6	37.9	13.5 → 9.8
Mask R-CNN	X101-FPN	42.8	43.9	38.4	39.7	6.8 → 5.5
Mask R-CNN	ConvNeXt-T-FPN	46.2	47.5	41.7	42.8	14.5 → 10.8
Mask R-CNN	Swin-T	46.0	47.5	41.7	42.9	9.8 → 7.4
Mask R-CNN	Swin-S	48.2	49.4	43.2	44.5	9.3 → 6.3
Mask R-CNN	ViT-B	51.5	52.0	45.7	46.8	3.9 → 3.3
Mask R-CNN	ConvNeXt-v2-FPN	52.9	53.4	46.4	47.6	5.4 → 5.2
HTC	R101-FPN	44.8	45.1	39.6	40.5	5.2 → 4.3
RTMDet-l	CSPX-PAFPN	51.1	51.4	43.7	44.0	3.4 → 2.5
RTMDet-x	CSPX-PAFPN	52.4	52.7	44.6	44.9	2.8 → 2.2
QueryInst	R101-FPN	49.0	49.8	42.9	44.1	3.5 → 2.5

These results show that the feedback strategy is not tied to a particular detector, mask head, or backbone. As expected, the extra refinement stages trade inference speed for accuracy.

Compatibility with training-free methods

All variants below were freshly evaluated with the same public Mask R-CNN R50-FPN 2x checkpoint on all 5,000 COCO 2017 validation images. Neither Turbo-Inference nor Soft-NMS requires retraining.

Method	Turbo-Inference	Soft-NMS	bbox AP	bbox AP50	bbox AP75	segm AP	segm AP50	segm AP75
Original Mask R-CNN			39.993	59.556	43.577	35.175	56.346	37.719
Mask R-CNN + Soft-NMS		✓	40.618	59.600	44.620	35.423	56.371	38.126
Mask R-CNN + Turbo	✓		40.279	59.482	43.901	36.400	56.498	39.277
Mask R-CNN + Turbo + Soft-NMS	✓	✓	40.901	59.510	44.945	36.643	56.487	39.688

Complementary training-free refinement

Turbo-Inference and Soft-NMS improve different parts of the prediction pipeline. Soft-NMS adjusts scores from pairwise box overlaps, while Turbo-Inference uses instance-mask feedback to refine boxes, scores, and masks. Their gains are therefore complementary:

• Turbo-Inference alone improves bbox AP by +0.286 and segm AP by +1.225 over the original inference path.
• Adding Soft-NMS on top of Turbo-Inference contributes another +0.622 bbox AP and +0.243 segm AP.
• The combined training-free pipeline reaches 40.901 bbox AP and 36.643 segm AP, improving the original model by +0.908 and +1.468, respectively.

This suggests that Turbo-Inference can be combined with other compatible training-free refinement or post-processing techniques for further gains, without updating the pretrained model weights.

The checkpoint used for verification has SHA-256:

3e542a40ccc2952293c56bddc05e2002d681b9f6f20fde01f5908a5540b582b3

Turbo-Inference trades additional head computation for accuracy. The backbone is not recomputed, but the refined RoIs pass through the mask head again.

Installation

This repository is built on MMDetection 3.3.0. A typical installation is:

conda create -n turbo-inference python=3.8 -y
conda activate turbo-inference

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -U openmim
mim install "mmengine>=0.7.1,<1.0.0"
mim install "mmcv>=2.0.0rc4,<2.2.0"
pip install -v -e .

See the MMDetection installation guide if your CUDA or PyTorch version differs.

Checkpoints

Pretrained weights are not stored in this repository because most files exceed GitHub's 100 MB per-file limit. Download them from their original release sources.

The default Mask R-CNN R50-FPN checkpoint used for the verified results is:

Model	Config	Checkpoint	SHA-256
Mask R-CNN R50-FPN 2x	config	download	3e542a40…b582b3

mkdir -p checkpoints
wget -P checkpoints \
  https://download.openmmlab.com/mmdetection/v2.0/mask_rcnn/mask_rcnn_r50_fpn_2x_coco/mask_rcnn_r50_fpn_2x_coco_bbox_mAP-0.392__segm_mAP-0.354_20200505_003907-3e542a40.pth

Complete checkpoint zoo used in our experiments

Model	Config / source	Official checkpoint
Cascade Mask R-CNN ConvNeXt-S	config	download
Cascade Mask R-CNN ConvNeXt-T	config	download
HTC R101-FPN 20e	config	download
HTC R50-FPN 1x	config	download
Mask R-CNN ConvNeXt-v2-B	config	download
Mask R-CNN ConvNeXt-T	config	download
Mask R-CNN R101-FPN 2x	config	download
Mask R-CNN R50-FPN 2x	config	download
Mask R-CNN Swin-S	config	download
Mask R-CNN Swin-T	config	download
Mask R-CNN X101-64x4d-FPN	config	download
QueryInst R101-FPN, 300 proposals	config	download
QueryInst R101-FPN, 100 proposals	config	download
QueryInst R50-FPN 1x	config	download
QueryInst R50-FPN, 300 proposals	config	download
QueryInst R50-FPN, 100 proposals	config	download
RTMDet-Tiny	config	download
ViTDet Mask R-CNN ViT-B	config	download

Quick Start

Download the default checkpoint above, then run:

python demo/image_demo.py \
  demo/demo.jpg \
  configs/mask_rcnn/mask-rcnn_r50_fpn_2x_coco.py \
  --weights checkpoints/mask_rcnn_r50_fpn_2x_coco_bbox_mAP-0.392__segm_mAP-0.354_20200505_003907-3e542a40.pth \
  --device cuda:0 \
  --out-dir outputs/turbo_mask_rcnn

Evaluate on COCO with one GPU:

python tools/test.py \
  configs/mask_rcnn/mask-rcnn_r50_fpn_2x_coco.py \
  /path/to/mask_rcnn_r50_fpn_2x_coco.pth

Or reproduce the eight-GPU regression:

./tools/dist_test.sh \
  configs/mask_rcnn/mask-rcnn_r50_fpn_2x_coco.py \
  /path/to/mask_rcnn_r50_fpn_2x_coco.pth 8 \
  --work-dir work_dirs/turbo_mask_rcnn_r50_8gpu

Prepare COCO in the standard MMDetection layout before evaluation:

data/coco/
├── annotations/instances_val2017.json
└── val2017/

Configuration

MaskBoxRefiner is registered in mmdet/models/utils/mask_box_refiner.py and is enabled from the RoI head:

roi_head=dict(
    type='StandardRoIHead',
    mask_box_refiner=dict(
        type='MaskBoxRefiner',
        box_threshold=0.20,
        score_threshold=0.35,
        mask_score_weight=0.35,
        empty_mask_fallback=False))

Option	Meaning
box_threshold	Foreground threshold used to derive a tight box from a mask.
score_threshold	Foreground threshold used when computing maskness.
mask_score_weight	Weight of the mask-aware score in score fusion.
empty_mask_fallback	Whether an empty foreground mask falls back to the input box.

Set mask_box_refiner=None to recover the upstream MMDetection inference path. The provided Mask R-CNN R50-FPN configuration also uses Soft-NMS with an IoU threshold of 0.5.

Stage-wise Refinement

For Mask R-CNN and HTC, four stages are used by default: detection, vanilla segmentation, turbo detection, and turbo segmentation. QueryInst uses three stages because another mask-head pass can disturb the interaction between its proposal features and refined RoI features.

Implementation Status

Turbo-Inference is fully implemented for every model family listed below. These paths can be used directly with existing pretrained weights and require no retraining.

Model family	Turbo-Inference
Mask R-CNN	✓
Cascade Mask R-CNN / HTC	✓
QueryInst / Sparse R-CNN	✓
RTMDet-Ins	✓

Roadmap

• Training-free Turbo-Inference for Mask R-CNN.
• Training-free Turbo-Inference for Cascade Mask R-CNN / HTC.
• Training-free Turbo-Inference for QueryInst / Sparse R-CNN.
• Training-free Turbo-Inference for RTMDet-Ins.
• Compatibility with Soft-NMS and other training-free refinements.
• Introduce the Turbo feedback loop during training and learn stronger Turbo-aware weights.

The current release focuses on training-free inference with existing checkpoints. The fresh Mask R-CNN Turbo-only and Turbo + Soft-NMS evaluation logs are stored in work_dirs/turbo_mask_rcnn_r50_no_softnms_8gpu/ and work_dirs/turbo_mask_rcnn_r50_readme_rerun_8gpu/, respectively. The remaining research direction is joint training with Turbo feedback to obtain better model weights. CondInst is intentionally excluded from this release.

Citation

If this project is useful in your research, please cite:

@article{zhao2026turboinference,
  title   = {A turbo-inference strategy for object detection and instance segmentation},
  author  = {Zhao, Zhen and Zhang, Gang and Hu, Xiaolin and Tang, Liang},
  journal = {Computer Vision and Image Understanding},
  volume  = {270},
  pages   = {104827},
  year    = {2026},
  doi     = {10.1016/j.cviu.2026.104827}
}

Preprint: arXiv:2606.12371.

Acknowledgements

This implementation is based on MMDetection. We thank the OpenMMLab contributors and the authors of the supported instance segmentation methods.

License

This codebase is released under the Apache 2.0 license. Please also follow the licenses of MMDetection, its dependencies, datasets, and pretrained models.