A computer-vision test of drone images reported higher scores for two segmentation systems when they were paired with a module that adds saliency and depth cues. The conditioned version of SAM 3, called SD-SAM 3, had the strongest overall results for instance segmentation, meaning the separate mapping of individual tower components. SD-Grounded-SAM had the highest precision and the strongest image-level semantic segmentation results among the methods tested.
The work is labeled an arXiv preprint, version 1, dated 26 August 2026. Its comparison covered 340 high-resolution images from three UAV flights around two communication towers, so the findings do not establish how the approach would perform on other tower configurations or independent infrastructure datasets.
A conditioning layer for cluttered views
The study paired original and conditioned versions of Grounded-SAM and SAM 3, producing four evaluated methods. Zero-shot means the pretrained systems were run in inference mode without updating their model parameters using the tower images. External box refinement was included only in the Grounded-SAM branch.
The added module combined appearance-based saliency, which highlights visually prominent regions, with monocular relative depth, an estimate of nearer and farther areas from a single image. Together, these signals were used to construct a foreground prior, or a guide to which parts of a cluttered scene should be treated as the target. The Grounded-SAM branch also used box refinement.
The images came from TOW-300, an author-constructed collection focused on tower-mounted equipment, primarily antennas and radio units. Forty images were used for inference-time hyperparameter selection, while 300 images were reserved for the final test.
Different systems favored different strengths
On the 300-image test set, SD-SAM 3 reported a mean average precision, or mAP, of 0.5701. Its reported AP values were 0.6270 and 0.6709, with recall of 0.7311, precision of 0.7817 and a mean matched intersection-over-union score of 0.9107. The paper reports this configuration as the strongest overall performer for instance-level segmentation.
SD-Grounded-SAM showed a different strength. It reported precision of 0.8970, semantic IoU of 0.6281 and Dice of 0.7716, the highest precision and strongest image-level semantic results among the evaluated methods.
Compared with unconditioned Grounded-SAM, the conditioned version reported mAP of 0.3780 versus 0.1737. AP@0.5 rose from 0.2236 to 0.4751, AP@0.75 from 0.1819 to 0.4129, recall from 0.3971 to 0.5073, precision from 0.4962 to 0.8970, and mean matched IoU from 0.8879 to 0.8989.
The corresponding comparison for SAM 3 showed mAP of 0.5701 with conditioning versus 0.4250 without it. AP@0.5 was 0.6709 versus 0.5438, AP@0.75 was 0.6270 versus 0.4831, recall was 0.7311 versus 0.6089, precision was 0.7817 versus 0.7662, and mean matched IoU was 0.9107 versus 0.8787.
The same pattern appeared in the aggregate semantic masks. SD-Grounded-SAM reported semantic IoU of 0.6281 and Dice of 0.7716, compared with 0.2876 and 0.4467 for Grounded-SAM. SD-SAM 3 reported 0.4376 and 0.6088, compared with 0.3630 and 0.5326 for unconditioned SAM 3.
The gains came with a tradeoff
A Grounded-SAM comparison with and without box refinement showed a shift between finding more candidates and rejecting false positives. With refinement, recall was 0.5073 versus 0.5226 without it, but precision was 0.8970 versus 0.5499. The refined version also reported higher mAP, 0.3780 versus 0.2569, and higher semantic IoU, 0.6281 versus 0.2920.
A separate SAM 3 comparison placed most of the reported change alongside the saliency step. mAP was 0.4250 without conditioning, 0.5681 with saliency alone and 0.5701 after depth was added. Recall followed a similar pattern, at 0.6089, 0.7281 and 0.7311. The extra depth cue corresponded to a smaller additional increase than saliency did in this comparison.
A narrow test leaves the larger question open
The evidence is limited to a comparison of pretrained models run in inference mode on TOW-300, with 40 images used for validation and 300 for testing. The authors identify evaluation across additional tower configurations and infrastructure datasets as future work, so the reported figures do not establish performance beyond this collection.
Paper data and sources
Original title: Saliency-Depth Conditioning for Zero-Shot Segmentation of Communication-Tower Components in Cluttered UAV Imagery
Authors: Ali Lesani, Chul Min Yeum, Su-Min Kang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text