AI Image Classification: Change the Backdrop and Keep the Product

Imagine a second-hand shop that photographs lamps against a cream wall and bags on a blue cloth. An image classifier suggests catalogue categories with impressive accuracy on the shop’s old photos. Then a volunteer puts a lamp on the blue cloth. The model calls it a bag. The mistake is useful: the shop now has a specific clue about what to test before letting the classifier sort a new batch of stock.

The question is whether the model recognises an object across the conditions the shop actually uses. A conventional accuracy score can tell the owner how often the model gets an answer right on a particular set of photos. It cannot, by itself, tell the owner what happens when the same object appears against a different backdrop. For that, you need a pair of images and a reason for changing just one thing.

What a background swap can reveal

In Noise or Signal: The Role of Image Backgrounds in Object Recognition, first posted in 2020 and presented at ICLR in 2021, Kai Xiao and colleagues separated foregrounds from backgrounds in images drawn from ImageNet. Among their tests, they compared images in which a foreground was placed on a background from the same class with images in which the background came from a randomly selected image. The researchers found that some of the models they examined could classify above chance from background information alone, and that changing the background could disrupt otherwise correct object predictions.

The authors’ Backgrounds Challenge shows how they constructed those tests. Their results establish that background cues mattered for the models and images they examined. They do not establish how a particular shop’s classifier will behave. The useful idea to borrow is the comparison: leave the target in view and change a cue that ought to make little difference to the answer.

Build pairs from photographs you could really receive

Start with objects that were not used to train or tune the classifier. Choose a few examples from each category you care about, including ones staff already find awkward to label. Photograph each object on its usual backdrop and on another plausible one. Keep the camera angle, distance and lighting as close as you can. If moving the backdrop also changes the shadows or makes the object harder to see, record that difference rather than treating the pair as a clean comparison.

Swap in both directions. Put a lamp on the blue cloth and a bag against the cream wall. If only the lamps change, you have learnt something narrower than ‘the classifier follows the backdrop’. Perhaps the blue cloth makes dark lamp bases hard to distinguish; perhaps one class has unusually varied examples. Photograph a few objects against a third, neutral surface to help you decide what to investigate next.

• Record the two labels for each object. Count correct labels that become wrong, and wrong labels that become correct. A single average can hide changes that cancel out.
• Take an empty shot of each backdrop. If the tool confidently offers a product category for an empty scene, record that separately. A classifier forced to choose from a fixed menu may return a label even when no valid choice exists. The blank photo tests how your workflow handles that case; it does not identify the model’s reasoning.
• Keep a separate set of ordinary new photos. Paired shots isolate a suspected weakness. Real incoming photos help you judge how often that weakness matters alongside other changes, such as framing, clutter and uneven light.

Decide which changes matter to the business

Suppose the same lamp gets the correct label on cream and the wrong one on blue. Repeat the pair on other lamps before changing the entire catalogue process. If several switch to ‘bag’ on blue and the reverse swap pulls some bags towards ‘lamp’, you have a pattern to act on. The test still cannot tell you precisely which pixels mattered, or whether the strongest cue was colour, fabric texture or the edge of the cloth. It can show that these plausible changes to the input altered decisions the shop wanted to stay stable.

A wrong suggestion is cheaper to catch when a volunteer already reviews each listing than when an automatic rule sends the item into the wrong collection. For the first workflow, the shop might keep suggestions while checking the affected classes. For the second, it might suspend automatic sorting for those classes, collect examples of each product against several realistic backgrounds, and repeat the test on fresh objects after making changes. If the supplier owns the model, give them the pairs and the expected labels. ‘The same lamp changes category when we move it’ is more actionable than a complaint about the average score.

Give context a clear job

Sometimes the surroundings are part of what the reader or customer wants to know. A marketplace could deliberately offer a search for products pictured in a nursery, where the room is relevant. But if the task is to assign an item to ‘lamp’ or ‘bag’, the room normally should not determine the label. The same choice appears in visual search for a specific catalogue item: a resemblance is useful only if it matches the purpose of the search.

Write one sentence before you run the test: ‘This label should stay the same when we change ___, provided ___ remains visible.’ Fill in the blanks with a change your team actually expects. Then show the owner three pairs and their results. If the model fails, you have a concrete condition to fix or route to review. If it passes, you have evidence about one change you can name. That makes a more useful release decision than another reassuring score with the backdrop left in place.

Sources

• Kai Xiao, Logan Engstrom, Andrew Ilyas and Aleksander Madry, Noise or Signal: The Role of Image Backgrounds in Object Recognition (preprint, 2020; ICLR, 2021).
• MadryLab, Backgrounds Challenge, the authors’ dataset and evaluation repository.