Source-linked AI summary
On Detection Reliability for Small Object Segmentation
Olivier Rukundo
TL;DR
This study asks whether detector-predicted boxes reliably indicate where small-object segmentation can be performed without risking missed segmentation outside the boxes. It matches adenovirus segmentation predictions to ground-truth objects and classifies their centroids relative to predicted boxes. The results show that matched predictions are not always inside detector boxes, so restricting segmentation to those regions risks excluding predictions and potentially introducing false negatives.
Problem
For small objects near one another or complex backgrounds, it is uncertain whether detector-predicted boxes can safely delimit segmentation without excluding objects a segmentor could identify.
Source: §1 · p. 1Method
The study matches predicted segmentation masks to ground-truth adenovirus objects using centroids, then records matches inside boxes, outside all boxes, or no matched segmentation.
Source: §1 · pp. 2–3Results
34 objects (47.89%) had matched segmentation centroids inside predicted boxes and eight (11.27%) outside for YOLO26x/YOLO26x-Sem; Mask R-CNN had 69 (97.18%) inside, none outside, and two (2.82%) unmatched.
Source: §3.3 · p. 6Takeaways & Limitations
Restricting segmentation to detector-predicted regions risks excluding matched segmentation predictions and potentially introducing false negatives.
Source: §5 · p. 7
Abstract
from arXiv · showhide
This study investigates whether predicted bounding boxes are reliable for determining where small-object segmentation should be performed without risking false negatives outside those boxes. The investigation leverages adenovirus particles as small objects, YOLO26x for detection, YO-LO26x-Sem for segmentation, and Mask R-CNN for both detection and segmentation. Segmenta-tion-model predictions were matched to ground-truth adenovirus masks or reference objects using their centroids, with ground-truth mask centroids preferred over ground-truth bounding-box centroids as references for matching. Matched predicted-mask centroids were considered as inside at least one detector-predicted bounding box or outside all such boxes, while ground-truth masks without matched predictions were recorded as having no matched segmentation. For the YOLO26 models, 34 reference objects (47.89%) had matched segmentation centroids inside at least one pre-dicted box, 8 (11.27%) had centroids outside all predicted boxes, and 29 (40.85%) had no matched predicted segmentation. For the Mask R-CNN model, the corresponding counts were 69 (97.18%), 0 (0.00%), and 2 (2.82%). These preliminary experiments demonstrated that restricting segmentation to detector-predicted bounding boxes risks excluding matched segmentation predictions outside those boxes, potentially introducing false negatives.
1. Introduction
Small-object segmentation can confuse nearby objects with noisy backgrounds, and detector boxes may omit objects a segmentor could otherwise identify. This study quantifies how predicted masks correspond spatially to detector boxes to assess the risk of restricting segmentation to those regions.
1. Introduction
- Segmentation can include foreground beyond annotated adenovirus boundaries, while detector-dependent segmentation can miss objects the detector overlooks.The YOLO26s-Sem and YOLO26m-Sem examples show foreground extending beyond ground-truth masks; a detector miss can prevent a box-prompted segmentor from producing a mask.Source: §1 · pp. 1–2, Figure 1 · p. 2
- Prior work includes bounding-box robustness evaluation, detector-error propagation, prompt sensitivity, and approaches for recovering segmentation beyond detector-proposed regions.The cited examples include a probabilistic framework and a soft-constraint rectangle approach.Source: §1 · pp. 1–2
- The study classifies matched predicted-mask centroids as inside or outside detector boxes and records ground-truth objects without a matched segmentation separately.This measures spatial correspondence between masks and boxes rather than relying on detection precision or an overall segmentation score alone.Source: §1 · pp. 2–3
- The experiments use adenovirus patches from the SSP dataset and compare detection and segmentation models in a TEM setting.The investigation connects prior adenovirus TEM studies by evaluating models on an SSP dataset derived from the adenovirus class of the TEM Virus Dataset.Source: §1 · p. 3
2. Materials and Methods
The methods assess correspondence between reference boxes and masks, specify the models and SSP data, and describe augmentation and training configurations. Reference objects are matched within images using centroid distances.
2.1. Reference Annotation Consistency Assessment
- Reference annotation consistency was checked by identifying mask components, expanding their boxes by five pixels, and matching box centres to mask centroids within each image.The diagnostic centre-coordinate agreement uses a five-pixel tolerance, distinct from the five-pixel box-expansion margin; mask-centroid containment is separately defined relative to reference boxes.Source: §2.1 · pp. 3–4
2.2. Detection and Segmentation Models
- The evaluated models include five YOLO26 detectors, five corresponding YOLO26-Sem segmentors, and Mask R-CNN; centroid containment uses YOLO26x, YOLO26x-Sem, and Mask R-CNN.Detection-confidence and foreground-probability thresholds were 0.25 and 0.50, respectively.Source: §2.2 · p. 4
2.3. Training datasets
- The SSP dataset contains 145 adenovirus image-patch pairs, split into 87 training, 26 validation, and 32 test patches, with 71 reference objects in the test split.Detection and segmentation share the image splits; reference objects are eight-connected foreground components.Source: §2.3 · p. 4
2.4. Augmentation operations
- Training-set augmentation uses random rotations, translations, scaling, and independent horizontal and vertical flips.Rotations range from −180° to 180°, translations reach 10% of image dimensions, and scaling ranges from 0.8 to 1.2.Source: §2.4 · p. 4
2.5. Hyperparameter settings
- Training allowed up to 300 epochs with early-stopping patience of 100, while batch sizes and optimizers differed across YOLO26, YOLO26-Sem, and Mask R-CNN.Mask R-CNN used SGD with an initial learning rate of 0.005, momentum 0.9, weight decay 0.0005, and cosine scheduling; manual adjustment stopped when a monitored metric reached 0.90 within 30 epochs.Source: §2.5 · pp. 4–5
3. Results
The results check reference-annotation consistency, evaluate YOLO26 variants, and compare predicted-mask centroid locations with detector boxes. Matched centroids sometimes fall outside predicted boxes, while qualitative examples distinguish those cases from objects without matched segmentation.
3.1. Reference Annotation Consistency Results
- Across 32 test patches and 71 reference objects, 69 center–mask pairs (97.18%) were within 5 pixels; all 71 mask centroids (100%) were inside their reference boxes.Two center–mask pairs exceeded the 5-pixel tolerance.Source: Figure 3 · p. 5, Figure 4 · p. 5, §3.1 · p. 5
3.2. Detection and Segmentation Performance
- All five YOLO26 detector and segmentor sizes were evaluated; YOLO26x and YOLO26x-Sem were selected for subsequent analysis because they generally led their respective model groups.The results include detection and segmentation performance tables for the SSP test split.Source: §3.2 · p. 5, Table 1 · p. 6, Table 2 · p. 6
3.3. Predicted-Mask Centroid Containment
- 34 objects (47.89%) had YOLO26 matched centroids inside predicted boxes, 8 (11.27%) outside; Mask R-CNN had 69 (97.18%) inside, 0 outside, and 2 (2.82%) unmatched.The remaining 29 YOLO26 objects (40.85%) had no matched predicted segmentation.Source: Figure 5 · p. 6, §3.3 · p. 6
3.4. Qualitative Examination of Predictions
- Figure 6 illustrates inside-box matches, outside-box matches, and ground-truth objects without matched segmentation; the last category does not mean no segmentation exists.Predictions may exist but fail the matching criterion for those ground-truth objects.Source: §3.4 · pp. 6–7, Figure 6 · p. 7
4. Discussion
The discussion interprets the centroid-containment results as a preliminary test of spatial correspondence, not a direct evaluation of segmentation rerun within detected boxes. It also identifies limits on generalizing the findings.
4. Discussion
- All 69 matched Mask R-CNN mask centroids fell inside predicted boxes, but this is constrained by associated box–mask outputs and does not establish detection completeness.The two ground-truth objects without matched segmentation indicate a matching outcome, not necessarily a detection failure.Source: §4 · p. 7
- The analysis tests whether matched predicted-mask centroids fall outside detector boxes, rather than evaluating segmentation rerun within those boxes..Source: §4 · p. 7
- Matching-sensitivity analysis and evaluation on broader datasets are needed before the findings can be generalized.Source: §4 · p. 7
5. Conclusions
The conclusion reports preliminary evidence that matched predicted-mask centroids are not always inside detector-predicted boxes. Restricting segmentation to those regions therefore risks excluding matched predictions and potentially introducing false negatives.
5. Conclusions
- Matched predicted-mask centroids were not consistently located inside detector-predicted bounding boxes.Source: §5 · p. 7
- Restricting segmentation to detector-predicted regions risks excluding matched segmentation predictions.Source: §5 · p. 7
- Such exclusions may potentially introduce false negatives, according to these preliminary findings.Source: §5 · p. 7