acceptodds
Under review as a conference paper at ICLR 2027

PANDA: Training-Time Privileged Attribute Supervision for Ship Re-Identification

Abstract

Ship Re-Identification requires distinguishing similar vessels across cameras. Identity labels tell the model which images match, but not which visual details matter. Language can provide this guidance, yet sentence-level alignment does not explicitly teach attribute discrimination, while MLLM-based reranking adds inference cost. To address these problems, we propose **P**rivileged **A**ttributes for **N**ative **D**escriptor **A**daptation (**PANDA**), a three-module framework that only uses language supervision during training. Specifically, **P**rivileged **A**ttribute **C**onstruction (**PAC**) extracts shared attributes from paired training views, organizes them into structured descriptions, and creates candidate negatives by replacing one slot with text from another identity. Using these annotations, **F**rozen-LM **A**ttribute **S**upervision (**FAS**) trains visual adapters to generate descriptions and distinguish matching attributes from alternatives through a frozen language model. **A**ttribute-**A**ware **F**usion (**AAF**) then combines identity and semantic features from the attribute-supervised encoder into a single retrieval descriptor. At inference, only the visual encoder and fusion head remain, requiring neither text nor a language model. Experiments on ShipReID-2400 and VesselReID show that PANDA outperforms the compared baselines while maintaining efficient inference, and ablation studies demonstrate the effectiveness of its components.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Ship ReID, Vessel ReID, Large Multi-modal Model, Multi-modal Large Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BXu6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a PANDA method for ship re-identification, which only uses ship attributes as privileged information during the training stage. A frozen MLLM extracts stable attributes from cross-camera pairs of the same ship and constructs hard negative samples by replacing one of the attribute slots with text from similar ships. Subsequently, the frozen language model is supervised by attribute generation (GAS) and slot discrimination (HSD) for the visual LoRA adapter; during inference, all language parts are removed and only visual descriptors are used. On the ShipReID-2400 and VesselReID datasets, PANDA outperforms the baseline methods based on visual and language enhancement in both module ablation experiments and annotation quality analysis.

Critical strengths

1、The main idea is good: encode fine-grained attribute information into the visual encoder during training, so that at inference only the visual encoder and a fusion head are needed, without any language model.

2、 PANDA gets solid results on both benchmarks (51.19 mAP on ShipReID-2400 and 76.02 mAP on VesselReID), which shows this training-time supervision can really improve the image-only descriptor.

Critical weaknesses

1、Fine-grained attribute quality is doubtful: many slots are class-level cues shared by same-type ships. HSD brings only 0.15 mAP gain on ShipReID-2400, and the annotations are only audited by another LLM, without human verification.

2、The method is traditional training-time supervision, not a new paradigm. No comparison against directly generating structured VLM descriptions and matching them at retrieval, so the claim that this route beats language-based retrieval is unverified.

Questions for authors

1、Many slots seem to be class-level attributes shared by same-type ships. Can you report how often two identities have identical slot contents, i.e., the real identity-level discriminability of the PAC attributes?

2、The annotation audit is done by another LLM without human check. Can you verify on a sampled subset that the fine-grained cues (window counts, equipment placement) are actually visible and correct in the images?

3、You only compare against shortlist rerankers. Could you add full-gallery retrieval with structured descriptions — generate PAC-style attributes per image and match them directly — compared on the same accuracy-latency curve?

4、It is unclear which frozen LM is used for FAS in the default setting — §4.1, D.4 and Tables 6–7 are hard to reconcile. Please clarify, and report total training cost including annotation.

Reviewer_Tk1zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

PANDA is a training-time language-supervision framework for Ship ReID. It updates visual adapters through a frozen language model and requires no language inference at deployment. Experiments on ShipReID-2400 and VesselReID report improvements over the compared baselines, supported by controlled supervision comparisons and component ablations.

Critical strengths
  1. Controlled supervision baselines support FAS on both datasets. It demonstrates the claim that “neither direct attribute prediction nor holistic language generation alone fully captures the benefit of FAS”.

  2. Higher accuracy than MLLM reranking without language-model inference. In Table 1, PANDA outperforms MLLM rerankers. It retrieves with an image-only descriptor, so it avoids the high inference cost of MLLM reranking.

Critical weaknesses
  1. Some details are unspecified in the main body. For example, PAC selects “a similar identity as the donor”, but the main body does not specify how similarity is measured. It also does not explain how Pure Visual ReID differs from w/o FAS, although Table 2 reports different results for them. Also, Table 1 does not report the visual backbones of the compared methods.

  2. Some gains are small and not supported by repeated runs. On ShipReID-2400, w/o HSD is only 0.15 mAP below Full (51.04 versus 51.19), and removing Entropy Scaling (H) changes every metric by around 1 points on both datasets. Since no standard deviations or repeated runs are reported, these gaps may be within run-to-run variance and do not clearly support the claim that “the attention entropy H provides additional benefit”.

Questions for authors
  1. Section 3.3 states “we select a similar identity as the donor”. What do “Top-K Nearest” and “Intersection” in Figure 2(b) represent?
  2. For w/o HSD in Table 2, is the second stage skipped? If training continues, what loss weight and training duration are used?
  3. AAF is described as retaining identity and attribute evidence. How do G and G + S compare in the w/o FAS setting of Table 2?
Reviewer_PZvQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The author propose a Privileged Attributes for Native Descriptor Adaptation (PANDA) to use training-time language supervision to improve ship re-identification without requiring language-model inference at deployment. which comprises three modules. Privileged Attribute Construction (PAC), Frozen-LM Attribute Supervision(FAS) ,Attribute-Aware Fusion (AAF) 。The method achieves promising results on two ship ReID benchmarks.

Critical strengths

1、The substantial improvements are made on the two reid benchmarks. 2、Three specific modules are carefully designed and reasonably results are analayzed. 3、The motivation of this method is clear and easy to follow.

Critical weaknesses

1.Limited methodological novelty, and the method combines three existing techniques. The conceptual contribution beyond existing language-guided representation learning and knowledge distillation is not sufficiently clear. 2.Uses Qwen3VL as pretrained visual encoder, and the comparison methods are different pretraining model configurations. More fair comparisons are needed. 3.More validations are needed, the experimentals results are limited to two ship reid datasets. More experimental results are expected to be reported to analyze the vidences of three key components proposed by the authors.

Questions for authors

1、What fundamentally distinguishes PANDA from existing language-guided attribute distillation methods? 2、Can the authro provide the generalization of this method and further report more results ? 3、How sensitive is the method to be incorrect or ambiguous ?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.