A General Method for Detection and Segmentation of Terrestrial Arthropods in Images

Key results

1. 01

   ShownLarger YOLOv8 backbones scored somewhat higher on the in-domain validation split. Medium and small were nearly tied, and nano lost the most, especially on its worst-case subdatasets.

   Models were trained for 300 epochs on 5153 images and validated on 978 images from the same 23 subdatasets. AP50: 0.937 (L), 0.932 (M), 0.933 (S), 0.914 (N). AP50-95: 0.872, 0.862, 0.86, 0.821. A few subdatasets drove the gaps between sizes: ATO, BIS and AMA for precision, AMI and BDA for recall. · 3.1 Experiment 1: Backbone size comparison; L Experiment 1: Precision-Recall Curve and…

2. 02

   ShownLeaving one subdataset out of training lowered F1 on that subdataset, mostly through lost recall. The mean change in precision was not significant, but it varied widely between subdatasets.

   There were 23 leave-one-out medium models (50 epochs) plus a full-data control, each scored only on the held-out subdataset: δF1 ≈ −7.1% [−12.1%, −0.4%], δR ≈ −10.3% [−13.9%, −6.6%], δP ≈ 4.5% [−4.8%, 13.7%]. δ is a normalized relative change (the share of possible loss or gain), not a raw difference. Precision fell sharply for BIS and SIT and rose for AMT and CAO. · 2.5.2 Experiment 2: Leave-one-out…; 3.2 Experiment 2: Leave-one-out cross-validation; 2.4 Evaluation

3. 03

   ShownTen more epochs of training on the held-out subdataset alone closed the gap to a full-data control that was also fine-tuned, with a small gain in precision.

   Fine-tuned leave-one-out vs fine-tuned control: δF1 ≈ 0.9% [−1.3%, 6.7%], δR ≈ 0.0% [−2.4%, 4.5%], δP ≈ 5.1% [1.7%, 9.9%]. Intervals are bias-corrected and accelerated bootstrap intervals, with each subdataset weighted by the logarithm of its instance count. · 3.2 Experiment 2: Leave-one-out cross-validation; Figure legends

4. 04

   SuggestedPairwise leave-two-out training showed how the subdatasets help one another. Most form one mutually beneficial cluster. The citizen-science images (ATO), Collembola on soil (CAI) and a moth light-trap set (DPS) sit apart, which points to domain gaps.

   136 + 1 medium models (20 epochs) on 16 of 23 subdatasets, scored by two-way redundancy ρ² on F1. 11 of 16 subdatasets fall in a central synergistic cluster. BIS and PIN form a separate pair that benefit each other. AMI-AMT (the same camera-trap system) has the lowest ρ². In a minimum spanning tree on a PCoA of the redundancy matrix, 10 of 15 edges are negative, and 3 of the 5 positive edges touch ATO, DPS or CAI. · 2.5.3 Experiment 3: Leave-two-out…; 3.3 Experiment 3: Leave-two-out cross-validation; M Experiment 3: Leave-two-out cross-validation…

5. 05

   ClaimedThe authors argue that deduplicating on polygon IoU instead of box IoU separates tightly packed individuals better, and that a two-stage NMS makes it affordable. No ablation or timing is reported to support either point.

   Stage 1 groups predictions into connected components using bounding-box IoU at a 0.05 threshold. Stage 2 runs polygon NMS within each component, which cuts polygon intersections below O(N^2). The only support in the text is one example of disjoint polygons whose boxes overlap (Fig. S7). No comparison with SAHI or standard YOLOv8 NMS is quantified. · 2.3 Inference; 4 Discussion; Figure legends