A method that changes how much an image is augmented according to the image's estimated difficulty recorded higher reported segmentation scores than standard, non-adaptive training on Oxford-IIIT Pet. It had the best mean Intersection over Union (mIoU), a score for how closely predicted regions match labelled regions, for U-Net, DeepLabV3 and SegFormer-B0. On binary Pascal VOC, its results also led several comparisons, including foreground IoU, the reported overlap score for the foreground class, for all three architectures. The document is an arXiv preprint dated 26 Aug 2026.
How the allocation works
Called Difficulty-Aware Sample Allocation, or DASA, the approach combines prediction ambiguity, training loss, class rarity and boundary complexity. It maps those signals to a bounded, sample-specific augmentation strength, assigning stronger augmentation to samples judged more difficult. The paper describes DASA as architecture-agnostic and says it requires no changes to the network. In the reported combination, ambiguity and loss each received a weight of 0.35, rarity 0.20 and boundary complexity 0.10.
Three architectures, two task formats
The central question was whether augmentation should be allocated at the sample level according to difficulty rather than applied uniformly. The evaluation used Oxford-IIIT Pet trimap images in a three-class segmentation setting and Pascal VOC converted to binary foreground-background segmentation. The comparison covered U-Net, DeepLabV3 and SegFormer-B0, with standard training, strong uniform augmentation, random weighted sample-specific allocation, and single-factor strategies based only on loss, rarity or boundary complexity.
All images in the reported protocol were 128 by 128 pixels. Baseline training lasted 20 epochs, while adaptive methods ran for three rounds of three training epochs. The learning rate and weight decay were both 0.001. DASA used four Monte Carlo passes to estimate ambiguity, with batch sizes of 20 for U-Net and DeepLabV3 and 8 for SegFormer-B0.
Oxford-IIIT Pet showed the strongest pattern
Oxford-IIIT Pet produced the clearest pattern. DASA had the highest reported mIoU among the compared methods for each architecture. The paper also reported mean Dice, or mDice, as a second segmentation-overlap score. For U-Net, DASA's mIoU/mDice were 0.714/0.813, compared with 0.692/0.794 for baseline. For DeepLabV3, the corresponding figures were 0.740/0.831 with DASA and 0.633/0.742 with baseline. SegFormer-B0 recorded 0.730/0.825 with DASA versus 0.711/0.808 for baseline.
The result was not simply a case of more augmentation helping every model equally. On Oxford-IIIT Pet, strong uniform augmentation reduced performance for U-Net and SegFormer-B0 but improved it for DeepLabV3. That architecture-specific pattern is consistent with the paper's case for choosing augmentation at the sample level, although it remains a descriptive result from the reported comparison.
Pascal VOC was less uniform
Pascal VOC was more mixed. On the binary task, DASA had the highest mIoU for U-Net and SegFormer-B0, while its DeepLabV3 mIoU was effectively tied with the boundary-only strategy. DASA had the best DeepLabV3 mDice and the highest foreground IoU for all three architectures. Its mIoU/mDice/foreground IoU were 0.601/0.713/0.442 for U-Net, 0.634/0.731/0.482 for DeepLabV3, and 0.630/0.732/0.468 for SegFormer-B0.
The scores came with extra training time
Averaged over the three architectures, DASA reached 0.728 mIoU on Oxford-IIIT Pet and 0.622 on binary Pascal VOC. These figures are point estimates from a descriptive comparison, not a causal estimate. The supplied analysis reports no uncertainty estimates or replicate variability, so it does not show how consistently the same gaps would appear across repeated runs.
The performance figures came with higher reported runtime on Oxford-IIIT Pet. DASA took 13.09 minutes versus 5.74 for U-Net, 13.50 versus 5.56 for DeepLabV3, and 18.66 versus 8.45 for SegFormer-B0, compared with baseline. Pascal VOC runtime values and runtime variability were not reported. The authors identify the extra stochastic passes needed for ambiguity estimation as one of the method's limitations.
What the preprint does not settle
The evaluation remains narrow: it covered one three-class Oxford-IIIT Pet task and one binary Pascal VOC task, and it did not test larger-scale multi-class benchmarks. The authors also flag the fixed aggregation weights as a limitation or future-work issue. Those gaps leave open whether the pattern will hold when the dataset, task format or difficulty weighting changes.
Within the reported tests, the pattern favored multi-factor sample allocation: DASA led mIoU on Oxford-IIIT Pet for every architecture and led foreground IoU on binary Pascal VOC for all three, while taking longer to run. Whether that balance holds beyond the tested task formats remains an open question.
Paper data and sources
Original title: Difficulty-Aware Sample Allocation for Adaptive Data Augmentation in Semantic Segmentation
Authors: Olasimbo Ayodeji Arigbabu, Abimbola Ismail Arigbabu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text