Space to Depth Downsampling for Small Object Detection

This direction is merged into another direction: spatial resolution degradation, space-to-depth convolution. Its last version stays readable and citable here.

Version 1 of 126 papers38 conceptsfrozen 2026-10-04

In the architecture of a standard Convolutional Neural Network (CNN), downsampling is a necessary evil. To build deep representations and maintain computational efficiency, networks must reduce the spatial resolution of feature maps. Traditionally, this is achieved through strided convolutions or pooling layers. However, for tasks involving "tiny" targets—such as microscopic cancer spots, distant UAV-captured pests, or small defects on circuit boards—this reduction is often where critical information goes to die.

Recent research has converged on a more surgical approach to dimensionality reduction: replacing information-discarding operations with Space-to-Depth (SPD) transformations and context-preserving modules.

The Downsampling Dilemma

Standard downsampling works by applying a filter that skips pixels (striding) or aggregates them (pooling). While effective for identifying large objects like dogs or cars, these methods act as a low-pass filter that blurs or erases fine-grained details. As Zini Jian and colleagues note in their work on cancer diagnostics, traditional strided convolutions lose the pixel-level data required to identify microscopic features.

When an object occupies only a few pixels, a single stride-2 convolution can effectively "step over" the target, removing it from the feature map entirely. This problem is echoed across domains, from underwater robotics to forestry surveillance, where the "weak visual cues" of small targets are easily discarded as noise.

Lossless Reduction via Space to Depth

The primary alternative gaining traction is the Space-to-Depth (SPD) transformation. Instead of selecting or averaging pixels, SPD rearranges them. A $2 \times 2$ block of spatial pixels is "unrolled" into the channel dimension. For example, a single-channel $4 \times 4$ image becomes a four-channel $2 \times 2$ image.

This mechanism is "lossless" because every original pixel is preserved; they are simply moved to a different coordinate system. Aoduo Li and the team behind Forestpest-YOLO demonstrate that this preservation is essential for detecting minuscule, camouflaged pests that would otherwise be erased during downsampling. Similarly, Liqiang Wang utilized SPD-Conv to boost the detection of photovoltaic cell defects, achieving a 5.1% improvement in precision by ensuring microscopic anomalies were not discarded.

Preserving Context in Lightweight Models

The shift toward SPD is often paired with the principle of "delaying downsampling." This idea, popularized by Forrest Iandola and Song Han in SqueezeNet, suggests that maintaining larger activation maps later into the network preserves the spatial information necessary for accuracy in parameter-limited models.

In modern lightweight architectures, this is being refined through "context-preserving" downsampling. For instance:

  • Parallel Paths: Haiyang Qin and others developed infrared ship detection algorithms that use parallel downsampling paths (combining max pooling with $1 \times 1$ convolutions) to capture both dominant features and fine details.
  • Depth-wise Separable Striding: In iris segmentation, Songze Lei replaced standard pooling with strided depth-wise separable convolutions, which act as a more "surgical" method for reducing dimensionality while preserving boundary details.
  • Multi-unit Downsampling: The Fully Adaptive Transformer (FAT) by Qihang Fan and Huaibo Huang uses multiple small depth-wise convolution units to downsample, which captures global context without the aggressive information loss of large-stride convolutions.

Why It Works: The Mathematical Integrity of Features

The intuition behind these methods is that small object detection is a high-resolution task masquerading as a low-resolution one. When Yi Ke Yun introduced SelfReformer, they highlighted that standard interpolation (like bilinear) degrades the mathematical integrity of ground truth labels. By using space-to-depth transformations for labels as well as features, the network maintains a direct, pixel-perfect mapping between the input and the target.

This "pixel-to-pixel" relationship is more critical than network depth for high-precision tasks. As shown in YOLOv5_ES by Zhensong Li, replacing strided convolutions with SPD allowed the model to slash its parameter count by 55% while increasing accuracy. The model didn't need more layers; it simply needed to stop throwing away the data it already had.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_space_to_depth_downsampling_for_small_ob_5c5c4dff,
  title = {Space to Depth Downsampling for Small Object Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/space-to-depth-downsampling-for-small-object-detection/rdr_5c5c4dff35c443e692c5df7bd760a255},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_space_to_depth_downsampling_for_small_ob_5c5c4dff_v1,
  title = {Space to Depth Downsampling for Small Object Detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/space-to-depth-downsampling-for-small-object-detection/rdr_5c5c4dff35c443e692c5df7bd760a255/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}