- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper proposes EHiRe for detail-intensive e-commerce image-text retrieval with low-resolution inputs. It combines Detail Region Rearrangement (DRR), which rearranges local details with global context; Important Tokens Modulation (ITM), which uses high-resolution teacher attention and CoT guidance during training; and Relevant Embedding Selection (RES), which guides embedding updates through task and teacher rewards. It also introduces EComClutter, containing 94,481 query-image pairs across 35 categories with three relevance levels. Extensive experiments on EComClutter demonstrate the approach's effectiveness for detail-intensive product retrieval across different backbones and input resolutions. Experiments on MMEB-V2 further support its transferability to broader image-text retrieval tasks.
Critical strengths
1.The method addresses the loss of fine-grained information in low-resolution retrieval by combining local visual evidence with query-relevant representation learning. High-resolution teacher guidance and graded-relevance supervision encourage the model to retain details needed to distinguish partial from complete query matches.
2.Experiments across different backbones and input resolutions demonstrate the approach's effectiveness for detail-intensive retrieval. Results on MMEB-V2 also support its transfer to a broader range of retrieval tasks.
Critical weaknesses
- Unclear attribution of performance gains. The comparisons do not adequately account for the extra training and domain-specific fine-tuning used by EHiRe. Section 5.1 gives EHiRe four extra epochs after two SFT epochs, while Table 6 compares its 0.846 average with 0.783 for the model trained only with SFT. Table 4 compares GME Original at 0.719 with GME plus EHiRe at 0.809, although fine-tuned GME* already reaches 0.774 in Table 2. Without these controls, the experiments do not convincingly show that EHiRe offers an advantage over standard fine-tuning.
2.Unverified data separation and label reliability. The paper does not explain whether training and test sets share images or products, and it provides too little evidence about the reliability of fully relevant labels. Section 5.1 states that the queries in the two sets are different, but gives no image or product overlap statistics. Section 3.2 reports that an average of 96% of annotators agreed with the MLLM consensus on 1,000 unanimous predictions, but gives no separate audit results for fully relevant pairs, which make up 9.27% of the dataset. Fully relevant labels determine Recall@K, and their reliability cannot be judged from the overall agreement rate. These gaps weaken confidence in the retrieval results, and the experiments do not provide convincing evidence of generalization to unseen products.
Questions for authors
-
Please clarify the novelty of EHiRe and its fundamental distinctions in design rationale from existing detail-aware retrieval, high-resolution teacher distillation, and relevance-supervised learning methods.
-
How many training epochs did each fine-tuned baseline in Table 2 receive? Are comparisons using the same total number of training epochs as EHiRe already available?
-
Do the training and test sets contain identical or nearly identical images, or images of the same products? If so, how much overlap is there?
-
In the 1,000 audited pairs, what are the sample counts and average annotator agreement with the MLLM consensus for each relevance level?
-
In Table 5, does DRR enlarge regions from the original images or from images already reduced to the stated resolution? How does this affect the interpretation of the claimed detail-preservation benefit?