Zone Evaluation: Revealing Spatial Bias in Object Detection

TL;DR

The paper identifies a pervasive spatial bias in object detectors, showing border regions underperform relative to the image center when evaluated with traditional AP. It introduces Zone Evaluation and Zone Precision ($ZP$) to quantify per-zone detection performance, revealing a substantial gap between inner and border zones across 10 detectors and 5 datasets. Through analyses of object scale, absolute position, and cross-zone data patterns, the authors argue that differences in zone-specific data distributions—not scale or position alone—drive this bias, motivating the concept of spatial disequilibrium. To address it, they propose Spatial Equilibrium Learning, comprising SELA (sampling-adjusted positive assignments) and SE Loss (loss weighting), which reduce $ZP$ variance and improve border-zone performance while largely preserving $AP$. The work provides extensive empirical evidence and offers practical paths toward more robust, zone-balanced object detection with broader implications for safety-critical applications.

Abstract

A fundamental limitation of object detectors is that they suffer from "spatial bias", and in particular perform less satisfactorily when detecting objects near image borders. For a long time, there has been a lack of effective ways to measure and identify spatial bias, and little is known about where it comes from and what degree it is. To this end, we present a new zone evaluation protocol, extending from the traditional evaluation to a more generalized one, which measures the detection performance over zones, yielding a series of Zone Precisions (ZPs). For the first time, we provide numerical results, showing that the object detectors perform quite unevenly across the zones. Surprisingly, the detector's performance in the 96% border zone of the image does not reach the AP value (Average Precision, commonly regarded as the average detection performance in the entire image zone). To better understand spatial bias, a series of heuristic experiments are conducted. Our investigation excludes two intuitive conjectures about spatial bias that the object scale and the absolute positions of objects barely influence the spatial bias. We find that the key lies in the human-imperceptible divergence in data patterns between objects in different zones, thus eventually forming a visible performance gap between the zones. With these findings, we finally discuss a future direction for object detection, namely, spatial disequilibrium problem, aiming at pursuing a balanced detection ability over the entire image zone. By broadly evaluating 10 popular object detectors and 5 detection datasets, we shed light on the spatial bias of object detectors. We hope this work could raise a focus on detection robustness. The source codes, evaluation protocols, and tutorials are publicly available at https://github.com/Zzh-tju/ZoneEval.

Figures (15)

• Figure 1: The traditional evaluation measures the detection performance over the entire image zone, but it neglects the measurement of local zone and can hardly reflect the spatial bias. Our zone evaluation (ZP, the average precision constrained in the zone) compensates for these issues, indicating a large performance gap between zones. The results are reported by GFocal gfocal on the VOC 2007 test set voc.
• Figure 2: The detector is less sanguine in detecting border objects. The visualizations are reported by GFocal gfocal. Zoom in for a better view.
• Figure 3: Definition of evaluation zone when $n=5$.
• Figure 4: Mean ZP with various object scale ranges. One can see that for each range of the object scale $r$, the spatial bias is significant on the three object detectors.
• Figure 5: (a) The Sudoku-style dataset is constructed by regularly placing all the objects of the test set on a $600\times600$ black image. (b) Zone evaluation on GFocal ($3\times3$ grids).