Brief Review — Deep-Recursive Residual Network for Image Semantic Segmentation

Using Recursive Blocks for Semantic Segmentation

Sik-Ho Tsang
3 min readFeb 24


Deep-Recursive Residual Network for Image Semantic Segmentation,
Zhang JNCA’20, by Donghua University, and City University of Hong Kong,
2020 J. Neural Computing and Applications, Over 15 Citations (Sik-Ho Tsang @ Medium)
Semantic Segmentation

1.5. Semantic Segmentation / Scene Parsing / Instance Segmentation / Panoptic Segmentation
2014 … 2021
[PVT, PVTv1] [SETR] [Trans10K-v2, Trans2Seg] [Copy-Paste] 2022 [PVTv2] [YOLACT++]
==== My Other Paper Readings Are Also Over Here ====

  • As recursive block can improve performance without introducing new parameters, recursive block is utilized with the concatenation layer to combine the output maps of the recursive convolution layers with same resolution but different field-of-views.


  1. Recursive Block Utilization
  2. Results

1. Recursive Block Utilization

1.1. Recursive Block

Recursive Block
  • Recursive block stacking three recursions with residual unit. If it is unfolded, this structure produces 8 paths for data to choose.
Illustration of Applying Recursive Blocks
  • Inspired by FPN, all output maps from recursive convolution layers are concatenated. That is to combine large field-of-view, semantically strong features with small field-of-view, semantically weak features via concatenation operation and the concatenated maps contain rich semantic information but with high channel dimension.
  • For the trade-off between computation and accuracy, the maps need to perform a 1×1 convolution layers to reduce channel dimension as well as merge all information.

1.2. Networks

  • Simplified FCN, and Simplified FCN with atrous convolution are used.
  • They are using VGG backbone. The fifth block which contains 3 convolution is replaced by the proposed recursive block.
The modified architecture of Mask R-CNN applying method.
  • Mask R-CNN is also modified by applying the proposed recursive block at the mask head.

2. Results

2.1. DeepFashion

DeepFashion Results
  • Although FCN-8s has little better results, simplified FCNs have fewer parameters while maintaining similar performance.

All 3 networks, after using proposed recursive blocks, the performance is maintained or even better.

Visualizations left to right is image, ground truth, original network, the proposed method

2.2. Cityscapes

Cityscapes Results

Similar trends for Cityscapes.

2.3. PASCAL VOC 2012

PASCAL VOC 2012 Results

Similar trends for PASCAL VOC 2012.

2.4. MS COCO

MS COCO Results

Mask R-CNN is used. Using recursive blocks obtains higher APs.



Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: for Twitter, LinkedIn, etc.