A computer-vision preprint describes a training method in which several different object detectors share knowledge step by step. In the reported MS COCO 2017 benchmark, the final carrier recorded 32.8 AP from a RetinaNet start and 33.2 AP from a Faster R-CNN start across the full 80-category support. Each result was 2.6 AP above its matched Avg-FPN knowledge-distillation control.
The study frames the problem around directed transfer difficulty: knowledge may be easier to move between some detectors than others, depending on the direction of transfer. It asks how that difficulty can determine a consolidation order, and whether the combined knowledge can then be returned to individual models without wiping out their specialist strengths.
The route is chosen by transfer difficulty
The experiment used MS COCO 2017, with train2017 for training and val2017 for evaluation. Four heterogeneous expert detectors were involved. The starting carrier supported 40 categories, each expert contributed 10 disjoint expert-specific categories, and the combined support covered 80 categories. The carrier began as either RetinaNet or Faster R-CNN.
The method, called TGRD, estimates operational transfer difficulty from held-out feature-alignment residuals, or the remaining mismatch when one model's internal features are aligned with another's. It then fixes those scores in a directed table and uses a greedy procedure to choose which expert should be incorporated next. Consolidation proceeds toward a carrier covering the union of categories before the accumulated knowledge is sent back to the individual experts.
The selected order depended on the carrier used at the start. From RetinaNet, the reported path was t1 to t4 to t3 to t2. From Faster R-CNN, it was t3 to t1 to t4 to t2. Expert t2 was selected last in both paths.
The comparison was made with simultaneous Avg-FPN knowledge distillation under the same detector society, data, annotations and evaluation protocol. Both approaches were given a matched 48-epoch training budget.
Coverage expands as original skills hold
The stepwise comparisons favored the final carrier in all four reported increments. For the RetinaNet initialization, the gains ranged from 1.5 to 6.5 AP, with a mean gain of 3.6 AP. For Faster R-CNN, gains ranged from 2.6 to 10.2 AP, averaging 5.9 AP.
Performance on categories already supported by the starting carrier remained close to its initial level. With RetinaNet, AP on those categories moved from 37.3 to 39.5, compared with 37.9 for the control. With Faster R-CNN, the final value was 39.1, within 0.9 AP of the starting carrier and 0.5 AP of the control.
The second phase reversed the direction of transfer, sending the consolidated carrier's knowledge back to the experts. Across eight expert-carrier evaluation pairs, the resulting reciprocal detectors reached 20.8 to 28.4 AP on categories they had not previously supported. Across the full 80-category union, they reached between 31.2 and 35.0 AP.
The reported changes on original supports ranged from minus 0.8 to plus 0.4 AP, with a mean absolute change of 0.28 AP. On expert-specific supports, every reciprocal detector stayed within 1.3 AP of its original expert, with a mean absolute change of 0.58 AP.
The paper also gives a conditional theoretical comparison. Under its stated assumptions, the proxy certificate for the progressive approach was no larger than the corresponding certificate for an aggregated target.
A benchmark result with a defined scope
The authors interpret the findings as support for order-aware consolidation followed by reciprocal transfer as a viable route to broader society-wide coverage while retaining detector specialization. The reported evidence comes from the specified MS COCO 2017 arrangement: four expert detectors, a 40-category starting carrier, an 80-category union and two starting architectures.
The document is an arXiv version 1 preprint dated 26 August 2026.
Paper data and sources
Original title: Socialized Detector Learning: Trajectory-Guided and Reciprocal Distillation for Heterogeneous Object Detectors
Authors: Weihao Li, Yunqi Zhu, Zhihe Fan et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text