PRISM-Net splits product retrieval into line and color routes, lifting SwinV2 line mAP@10 by 5.74 points

Product retrieval no longer relies on a single holistic embedding; structural contour and chromatic appearance go to two dedicated expert groups, with a global category pathway kept alongside. Holistic CNN and vision Transformer features mix background illumination or dominant color blocks with category identity, so a metallic tool query often returns specular kitchenware and a crimson garment query returns clothing of the wrong color. On the SwinV2 backbone, the line-only configuration raises line mAP@10 from 56.53% to 62.27% and the color-only configuration raises color mAP@10 from 26.06% to 35.18%; PRISM-Net itself reaches 83.42%, 57.08% and 31.43% on product, line and color mAP@10.

The line pathway holds 3 residual experts and the color pathway 10, and each router emits independent Sigmoid probabilities, so one product can carry several color tones or mixed contour scales. Conventional mixture-of-experts routing is optimized mainly by downstream task loss and load balancing, which neither guarantees line-color separation nor prevents experts from becoming redundant feature extractors. Frozen SigLIP-base-patch16-224 turns 3 line prompts and 10 color prompts into soft targets distilled to the routers by point-wise binary cross-entropy, with stop-gradient detaching router parameters from classification gradients.

A meta-router estimates a gate between 0 and 1 per image, convexly combines the line and color features, and concatenates the result with the global base feature into the final retrieval representation. Fixed concatenation or globally shared expert coefficients cannot reflect how differently categories depend on contour versus appearance. Across 55,091 test images, DINOv3 assigns Appliances 6.89 times the line weight of Women's Clothing, and apparel puts more than 96% of its weight on the color expert, without manual attribute labels.

The demonstrated scope is the 18 categories and 275,454 high-resolution images retained after cleaning the Amazon Products dataset, under 5-fold cross-validation with one fold of 55,091 queries against a 220,363-image gallery, evaluated on the SwinV2, ConvNeXt-V2 and DINOv3 backbones. The intended readers are researchers and engineers working on multimedia retrieval representation learning.

Gains depend on matching backbone and routing configuration: SwinV2 improves clearly, ConvNeXt-V2 PRISM-Net product mAP@10 is 0.83 points below its baseline, and on DINOv3 the full model does not exceed the base micro- or macro-accuracy. The authors call the DINOv3 behavior teacher-student capacity inversion, since its native pixel-level edge sharpness exceeds the linguistic resolution of SigLIP's three prompts, and note that routing behavior should be read against category appearance and backbone inductive bias. The abstract and the retrieval results table disagree on which SwinV2 configuration owns the product and color mAP@10 figures: in the table, 83.90% and 35.18% belong to the color-only configuration, while PRISM-Net reaches 83.42% and 31.43%.