19Computer vision · Read 4 min

Image segmentation: isolating the product before computing similarity

Separating a product from its background belongs in the search engine only when results improve without excessive latency or image loss.

Article context

A pixel-level mask can reduce background influence, provided its effect on final retrieval is evaluated.

CONTEXT

This post expands on a path explored during my visual-search work: applying YOLOv8 instance segmentation to product-cluster prototypes before computing their embeddings. The hypothesis was straightforward: in a heterogeneous catalogue, backgrounds, models and art direction may influence the visual representation as strongly as the target product.

SUMMARY

Segmentation assigns a mask to the pixels belonging to each detected object. The product can then be cut out, cropped or recomposed on a neutral background before entering Contrastive Language–Image Pre-training (CLIP). The experiment compares full and segmented images across four dimensions: mask quality, catalogue coverage, neighbour relevance and computational cost.

Key diagram

Compare the full image and segmented product before indexing

01 · FULL IMAGE

Product + scenery + lighting

VS↔SAME ENCODER

02 · ISOLATED PRODUCT

MASK

Mask + controlled background

→INDEX

03 · NEIGHBOURS

The ranking decides

04 · ABLATION

RELEVANCE+ ?

COVERAGE? %

LATENCY+ ? ms

The diagram represents a simplified ablation: only preprocessing changes between branches. Neighbour relevance, coverage and latency determine whether the mask belongs in the pipeline.

A research-based explainer grounded in the sources listed below. Some mechanisms are simplified without changing the overall conclusions.

Experiment19

Does scenery influence the embedding more than the target product?

I applied YOLOv8-seg to visual-cluster prototypes to compare full images, crops and cut-out products before encoding. This ablation isolates the effect of preprocessing without changing the model or index.

Decision criteria

Neighbour relevance by category

Coverage and mask failures

End-to-end pipeline latency

Takeaways

Segmentation belongs in the pipeline only when it improves retrieval for the relevant categories without unduly reducing coverage or slowing processing.

A better mask does not guarantee better neighbors: evaluate the final ranking too.
Compare full images, crops and masks with the same catalogue and encoder.
Provide a fallback image and analyze errors by category and photography style.

01 — Understand the task

Segmentation produces a dense prediction at pixel level.

Classification assigns one category to a full image. Detection locates an object with a bounding box. Semantic segmentation assigns a class to every pixel, while instance segmentation also separates distinct objects from the same class. Two shoes in one scene therefore receive two different masks.

Fully Convolutional Networks (FCNs) established an architecture producing class maps at image resolution. Mask Region-based Convolutional Neural Network (Mask R-CNN) later added a mask branch to an object detector. YOLO instance-segmentation models pursue the same goal with an architecture designed for fast inference.

02 — Product hypothesis

Removing scenery can shift embeddings toward product shape and appearance.

An embedding model transforms the entire image into a vector. In fashion photography, that vector may encode the garment, model pose, studio, lighting and background. Such context is sometimes useful; it can also group products by art direction instead of cut, material or colour.

I therefore explored YOLOv8-seg on prototypes produced by clustering. The mask supports several preparations for comparison: full image, bounding-box crop, cut-out product on a neutral background and cut-out product with a margin of context.

03 — Protocol

Evaluate the mask and the retrieval engine within the same experiment.

Segmentation quality can be measured with Intersection over Union (IoU), the overlap between predicted and reference masks divided by the area of their union. Mean Average Precision (mAP) summarizes detection and mask quality across thresholds and categories. These metrics describe the model’s visual output.

The product decision requires a second set of measures: Recall@K, Precision@K or Normalized Discounted Cumulative Gain (NDCG) for neighbours, successful-segmentation coverage, per-category results, latency and cost. A highly accurate segmenter may still add little value when neighbours do not improve or too many catalogue items leave the pipeline.

04 — Decision

Use segmentation for categories where context demonstrably degrades retrieval.

The protocol is an ablation: the same catalogue, encoder and index are evaluated with and without masks. Differences are analyzed by category, photography style and segmentation confidence. This comparison identifies cases where scenery creates substantial unwanted variance.

In the project, segmentation remained in the experimental toolbox rather than entering the first service. Generic CLIP already supported raw visual similarity, while segmentation added another stage, possible failures and a cost that required justification. Final retrieval quality therefore remains the decision criterion.

05 — From concept to decision

Decide by photography style rather than for the whole catalogue.

A garment photographed in a busy setting may benefit from masking; styling may also reveal fit or use that helps retrieval. I would compare image preparations by photography style and retain the full image when segmentation fails.

The decisive test would examine retrieved neighbors and the proportion of products still served. Masks would be retained where their benefit outweighs lost context, failures and additional cost.

Primary source

YOLOv8-seg exploration for visual search

Experimental path developed in the Visual Search notebooks: segmenting cluster prototypes before embedding computation.

Selected references