When a retrieval system must describe 200 bird species with a 12-bit binary code, the code space is too small to keep every species equally far from every other. The Welch bound, a lower limit on how similar some pair of $C$ vectors in $d$ dimensions must be, forces some pairs of classes to sit close together. Fine-grained categories make this worse. Many species share the same overall body plan and differ only at one small spot, such as an eye ring, a wing bar, or the beak tip. Current low-bit hashing methods pool the whole image into content-only tokens before producing bits. As a result, most of the code budget describes the shared template, while the few pixels that separate the most-confused species are averaged away before any bit can depend on them.
We propose to tie each hash bit (or small group of bits) to a location on the object itself rather than to the whole image. The network keeps a higher-resolution feature map. From its own attention it estimates a cheap object-centered coordinate frame (center, scale, and principal axis). Each bit's attention then carries a learned spatial preference in that frame, for example "near the head, slightly forward." The testable claim is not simply that this improves accuracy. It makes a specific prediction: the gains should be largest exactly where the code is most crowded, meaning at the shortest code lengths and on the most-confused class pairs, and should shrink as bits become plentiful.
Core Claim
Under a tight bit budget, crowding in fine-grained codes has two sources. One is geometric and unavoidable: the Welch bound on how well $C$ prototypes can be spread in $d$ dimensions. The other is architectural and avoidable. Pooled features cannot represent the small, localized differences between near-identical classes, so the encoder places those classes nearer each other than the geometry requires, and pose changes spread each class across the code space. Binding bits to object-centered coordinates on a higher-resolution map attacks only the avoidable part. Measuring the gap between achieved separation and the Welch-limited ideal therefore gives a direct test of the claim.
Where This Sits
The closest work is QueryOpt, Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization. It frames fine-grained hashing as set prediction. Learnable queries cross-attend to backbone features, and each query governs a hash bit. It also identifies Welch-bound crowding, which it fights with a circular-shift auxiliary branch that raises the dimensionality only during training. QueryOpt motivates this proposal and supplies its codebase (SEU-VIPGroup/QueryOpt). Its queries, however, attend on content alone, with no positional anchor, and it addresses saturation purely inside the latent space. We keep its query-to-bit structure and change what the queries can see and where they may look.
Coordinate injection supplies the mechanism but also constrains it. CoordConv appends normalized pixel-coordinate channels to convolutions. It fixes coordinate-regression failures, yet adds only +0.04% on ImageNet classification. We read that result as a warning, not a dead end. Canvas coordinates are useless when objects move freely around the image, which is also the known weakness of canvas coordinates under changes in pose and scale. Learning Canonical 3D Object Representation for Fine-Grained Recognition shows the opposite case. It projects features onto a canonical 3D mesh indexed by periodic positional encodings, and object-centered coordinates then pay off (88.4% on CUB, 94.7% on Stanford-Cars). That method needs a template mesh per category and a differentiable renderer. Our object frame is a 2D moment estimate computed from attention, with no meshes and no part labels, and we test it in the low-bit setting that paper never examines.
High-resolution preservation supplies the substrate. Beyond Part Models removes the last stride of ResNet-50 and slices the enlarged map into six fixed horizontal stripes, which works only for upright pedestrians. Lite-HRNet keeps parallel high-resolution streams cheaply, but it collapses them by global pooling when used for classification. We use the stride-removal idea and replace fixed stripes with soft, learned spatial preferences in an object-centered frame.
Work on embedding geometry sets the lens for evaluation. Generalizing and Decoupling Neural Collapse via Hyperspherical Uniformity Gap and Hyperspherical Classification with Dynamic Label-to-Prototype Assignment spread class prototypes evenly when $d < C-1$, but both operate on pooled 1D vectors and test mostly on coarse datasets. Beyond Point-wise Neural Collapse reports that fine-grained CUB classes form curved, extended feature manifolds rather than points. This finding partly contradicts the point-prototype view. We address it directly: if pose drives those manifolds, object-centered coordinates should shrink them, and we measure whether they do.
Research Program
The backbone is ResNet-50 with the final stride removed, which roughly doubles the side length of the output map compared with the standard 7×7. Following QueryOpt, $K$ learnable queries each produce one or more bits.
From the mean attention over queries we compute an attention-weighted centroid $c$ and covariance $\Sigma$ of token positions. Each token position $p$ is then mapped to object coordinates $u = \Sigma^{-1/2}(p - c)$. The sign of each principal axis is fixed by the sign of the third moment along it, so the frame roughly accounts for translation, scale, and in-plane rotation without any supervision. Query $k$ attends to token $t$ with logit
$$a_{k,t} = \frac{q_k^\top x_t}{\sqrt{m}} + w_k^\top \phi(u_t),$$
where $x_t$ is the token feature, $m$ its dimension, $\phi$ a small set of low-frequency Fourier features of the object coordinate $u_t$, and $w_k$ a learned vector giving query $k$ its spatial preference. The second term is the only new ingredient. Setting $w_k = 0$ recovers content-only QueryOpt, and replacing $u_t$ with raw canvas coordinates recovers a CoordConv-style variant. Everything else, including losses, the circular-shift branch, and code binarization, is kept from the QueryOpt codebase.
The analysis side is equally central. For each trained model and code length, we compute class code centroids and report two quantities. The first is the largest pairwise coherence among centroids divided by the Welch bound, where a value near 1 means crowding is purely geometric. The second is the Hamming separation of the 50 most-confused class pairs. We also report a within-class spread score (the neural-collapse variability ratio), computed separately for CUB images binned by pose. Pose bins come from CUB's part annotations, which are used only for evaluation, never for training. The same annotations let us check whether each query's learned spatial preference lands on a consistent anatomical part across images.
Baselines
The naive reference point is a ResNet-50 at standard stride with global average pooling, followed by a linear hashing layer trained with the same retrieval loss. It shows how much any query structure helps at all.
The strongest existing method is QueryOpt, run from its released code. Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization reports 72.19% mAP on CUB-200-2011 with 12-bit codes, which it describes as a 15.77% improvement over AGMH. We will first reproduce this number under our hardware and seeds, and all comparisons will use the reproduced value. Our model shares its backbone, input resolution, epochs, optimizer, and a matched hyperparameter search of the same number of trials on the same validation split. Beating it means a mean mAP gain across three seeds that exceeds the seed spread at 12 and 24 bits on CUB, with the same direction on Aircraft and NABirds. For those two datasets the file gives no absolute baseline score: A Feature Consistency Driven Attention Erasing Network for Fine-Grained Image Retrieval reports only a relative gain of up to 29.78% over ExchNet at 32 bits on Aircraft, and no reported result was found on Lacuna for NABirds hashing. We will therefore also run FCAENet under the same protocol where its code allows.
Three ablations isolate the key ingredient. The first is QueryOpt on the stride-removed backbone with $w_k = 0$, which tests whether resolution alone accounts for the gains. The second uses canvas coordinates, which tests whether an object-centered frame matters or any positional signal suffices. The third adds a learned absolute positional embedding per token, the transformer-style alternative. The proposal succeeds only if the full model beats all three, and only if its margin over them grows as the bit budget shrinks.
Why not just…?
Why not just remove the stride and keep content-only queries? This is the most serious threat. A higher-resolution map may let content attention find eye rings and wing bars on its own, with no coordinate term. The first ablation tests exactly this. If it matches the full model at every code length, we will report that resolution, not coordinates, carries the effect, and we will drop the coordinate claim. The prior evidence cuts both ways. Beyond Part Models shows that resolution helps, but it also needed an explicit spatial partition to use that resolution.
Why not just add CoordConv channels? Canvas coordinates bind features to image positions, while birds and aircraft appear anywhere and at any orientation. The +0.04% ImageNet result in CoordConv is consistent with that failure. The canvas-coordinate ablation measures it directly, and the pose-binned spread score should separate the two variants: only object-centered coordinates should shrink within-class spread across pose bins.
Why not just use more bits, or train in a higher dimension? Retrieval deployments fix the code length, so larger inference codes are not an option. Training-time dimensional boosting is already QueryOpt's circular-shift branch, which we keep in every variant. The question is what that branch leaves unsolved. If our gains vanish once the branch is present, the avoidable-crowding hypothesis fails.
Why not just use a canonical 3D representation? It works well for recognition Learning Canonical 3D Object Representation for Fine-Grained Recognition, but it requires category template meshes and a renderer, which do not exist for many fine-grained domains. We treat its 88.4% CUB classification accuracy as the reference point for a secondary classification track. In that track, a 32-dimensional cosine classifier (so $d < C-1$) sits on our features. We expect to fall short of that number and report the gap honestly, while arguing for generality rather than peak accuracy.
Proposed Experiments and Decisive Checks
The main experiments are CUB-200-2011, FGVC Aircraft, and NABirds retrieval at 12, 24, 32, and 48 bits, with three seeds each. Stanford-Cars is added as a rigid-object contrast. For Cars retrieval, no reported hashing result was found on Lacuna, so all its baselines are run by us. A Lite-HRNet backbone serves as a second high-resolution substrate, to show that the effect is not specific to stride removal.
Three checks decide the proposal. First, the gain over the stride-removed, content-only ablation must be largest at 12 bits and shrink toward 48 bits. A flat gain across code lengths would indicate a generic capacity effect, not relief of crowding. Second, the coherence-to-Welch ratio and the separation of the 50 most-confused pairs must improve more than average-pair separation does, showing that the benefit lands where crowding bites. Third, within-class spread across CUB pose bins must fall under object-centered coordinates but not under canvas coordinates. Such a result would indicate that some of the extended class manifolds reported by Beyond Point-wise Neural Collapse come from pose rather than from intrinsic class structure. If all three hold, spatially anchored bits are a real remedy for the avoidable part of code crowding. If only the first ablation's resolution effect survives, the result is still a clean negative: in low-bit fine-grained hashing, preserving resolution matters and coordinates do not.
References
- Learning Attribute-Aware Hash Codes for Fine-Grained Image Retrieval via Query Optimization
- An intriguing failing of convolutional neural networks and the CoordConv solution
- Learning Canonical 3D Object Representation for Fine-Grained Recognition
- Beyond Part Models: Person Retrieval with Refined Part Pooling (and A Strong Convolutional Baseline)
- Lite-HRNet: A Lightweight High-Resolution Network
- Generalizing and Decoupling Neural Collapse via Hyperspherical Uniformity Gap
- Hyperspherical Classification with Dynamic Label-to-Prototype Assignment
- Beyond Point-wise Neural Collapse: A Topology-Aware Hierarchical Classifier for Class-Incremental Learning
- A feature consistency driven attention erasing network for fine-grained image retrieval