# CARAFE++: Unified Content-Aware ReAssembly of FEatures

arXiv: https://arxiv.org/abs/2012.04733 · HTML: https://mlsanity.com/paper/2012.04733/

## Key facts

- Cited by 20 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 7 Dec 2020.

## Built on by (most cited first, 20 of 20)

- [V3Det: Vast Vocabulary Visual Detection Dataset](https://mlsanity.com/paper/2304.03752/) (2023)
- [Frequency-aware Feature Fusion for Dense Image Prediction](https://mlsanity.com/paper/2408.12879/) (2024)
- [Few-shot Medical Image Segmentation with Cycle-resemblance Attention](https://mlsanity.com/paper/2212.03967/) (2022)
- [SAPA: Similarity-Aware Point Affiliation for Feature Upsampling](https://mlsanity.com/paper/2209.12866/) (2022)
- [FADE: Fusing the Assets of Decoder and Encoder for Task-Agnostic Upsampling](https://mlsanity.com/paper/2207.10392/) (2022)
- [UniNet: Unified Architecture Search with Convolution, Transformer, and MLP](https://mlsanity.com/paper/2110.04035/) (2021)
- [ELSA: Enhanced Local Self-Attention for Vision Transformer](https://mlsanity.com/paper/2112.12786/) (2021)
- [Learning to Aggregate Multi-Scale Context for Instance Segmentation in Remote Sensing Images](https://mlsanity.com/paper/2111.11057/) (2021)
- [A Refreshed Similarity-based Upsampler for Direct High-Ratio Feature Upsampling](https://mlsanity.com/paper/2407.02283/) (2024)
- [Butter: Frequency Consistency and Hierarchical Fusion for Autonomous Driving Object Detection](https://mlsanity.com/paper/2507.13373/) (2025)
- [FADE: A Task-Agnostic Upsampling Operator for Encoder-Decoder Architectures](https://mlsanity.com/paper/2407.13500/) (2024)
- [M2MRF: Many-to-Many Reassembly of Features for Tiny Lesion Segmentation in Fundus Images](https://mlsanity.com/paper/2111.00193/) (2021)
- [On Point Affiliation in Feature Upsampling](https://mlsanity.com/paper/2307.08198/) (2023)
- [Q2A: Querying Implicit Fully Continuous Feature Pyramid to Align Features for Medical Image Segmentation](https://mlsanity.com/paper/2404.09472/) (2024)
- [UniNet: Unified Architecture Search with Convolution, Transformer, and MLP](https://mlsanity.com/paper/2207.05420/) (2022)
- [Dual Complementary Dynamic Convolution for Image Recognition](https://mlsanity.com/paper/2211.06163/) (2022)
- [WCCNet: Wavelet-context Cooperative Network for Efficient Multispectral Pedestrian Detection](https://mlsanity.com/paper/2308.01042/) (2023)
- [Lighten CARAFE: Dynamic Lightweight Upsampling with Guided Reassemble Kernels](https://mlsanity.com/paper/2410.22139/) (2024)
- [A3-FPN: Asymptotic Content-Aware Pyramid Attention Network for Dense Visual Prediction](https://mlsanity.com/paper/2604.10210/) (2026)
- [ViT-Up: Faithful Feature Upsampling for Vision Transformers](https://mlsanity.com/paper/2606.14024/) (2026)

## Abstract (arXiv)

Feature reassembly, i.e. feature downsampling and upsampling, is a key operation in a number of modern convolutional network architectures, e.g., residual networks and feature pyramids. Its design is critical for dense prediction tasks such as object detection and semantic/instance segmentation. In this work, we propose unified Content-Aware ReAssembly of FEatures (CARAFE++), a universal, lightweight and highly effective operator to fulfill this goal. CARAFE++ has several appealing properties: (1) Unlike conventional methods such as pooling and interpolation that only exploit sub-pixel neighborhood, CARAFE++ aggregates contextual information within a large receptive field. (2) Instead of using a fixed kernel for all samples (e.g. convolution and deconvolution), CARAFE++ generates adaptive kernels on-the-fly to enable instance-specific content-aware handling. (3) CARAFE++ introduces little computational overhead and can be readily integrated into modern network architectures. We conduct comprehensive evaluations on standard benchmarks in object detection, instance/semantic segmentation and image inpainting. CARAFE++ shows consistent and substantial gains across all the tasks (2.5% APbox, 2.1% APmask, 1.94% mIoU, 1.35 dB respectively) with negligible computational overhead. It shows great potential to serve as a strong building block for modern deep networks.