.png){kind=icon}
返回
Consistencydet: few-step self-consistent box denoising for efficient object detection
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-026-04674-w.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
迭代框去噪 recently reformulated object detection as progressive noise-to-box refinement. However, diffusion-style detectors such as DiffusionDet inherit a key diffusion bottleneck: accurate inference often needs long reverse chains and repeated decoder calls. This paper proposes ConsistencyDet, a few-step self-consistent box denoising framework for closed-set object detection. Our motivation is to transfer the few-step sampling advantage of consistency models from generative modelling to discriminative detection. During training, Gaussian-corrupted ground-truth boxes and RoI features are decoded by a shared noise-conditioned detector. We introduce a detection-specific consistency objective along a probability-flow ordinary differential equation: adjacent noisy box states are decoded with shared weights and supervised by the same matched clean detection targets. At inference, random box proposals are refined in a few denoising steps, while Box-renewal replaces low-confidence proposals to keep the proposal distribution close to the training corruption process. On MS-COCO with ResNet-50, ConsistencyDet obtains 46.80 AP at two sampling steps and 6.892 FPS, exceeding DiffusionDet’s 46.54 AP at twenty steps while being about 7.2 times faster under the same RTX3080 protocol. Experiments on MS-COCO and LVIS with convolutional and transformer backbones show competitive accuracy with much lower long-chain sampling cost. Code, configurations, and reproduction scripts are released at https://github.com/Tankowa/ConsistencyDet .
Keyword:
Efficient object detection
Box denoising
Consistency models
Diffusion-style detection
Few-step sampling
Probability-flow ODE
Reproducible visual computing benchmark
期刊
IF:
2.9
论文数:
4.6K
被引数:
6.5K
机构
引用论文
.png){kind=icon}
没有更多内容
没有更多内容