ARC: Anchored Representation Clouds for High-Resolution INR Classification

TL;DR

ARC introduces Anchored Representation Clouds to inject locality into implicit neural representations by anchoring latent vectors in image coordinates and querying the nearest anchors with a shared decoder. This enables high-resolution INR classification and robust performance under image-space transformations, while permitting weight-space data augmentation and downstream point-cloud classification via PTv3. The approach achieves strong results on high-res datasets like Imagenette (full and cropped resolutions) and competitive or state-of-the-art performance on standard INR benchmarks, with demonstrated improvements in translation robustness and training efficiency. The work suggests practical pathways for scalable, memory-efficient INR-based classification and opens avenues for end-to-end integration of fitting and classification along with dynamic capacity adaptation.

Abstract

Implicit neural representations (INRs) encode signals in neural network weights as a memory-efficient representation, decoupling sampling resolution from the associated resource costs. Current INR image classification methods are demonstrated on low-resolution data and are sensitive to image-space transformations. We attribute these issues to the global, fully-connected MLP neural network architecture encoding of current INRs, which lack mechanisms for local representation: MLPs are sensitive to absolute image location and struggle with high-frequency details. We propose ARC: Anchored Representation Clouds, a novel INR architecture that explicitly anchors latent vectors locally in image-space. By introducing spatial structure to the latent vectors, ARC captures local image data which in our testing leads to state-of-the-art implicit image classification of both low- and high-resolution images and increased robustness against image-space translation. Code can be found at https://github.com/JLuij/anchored_representation_clouds.

Figures (10)

• Figure 1: ARC anchors latent vectors directly in image coordinate space, preserving the local spatial image structure within the INR weight-space. Once trained, ARC can be processed by a point cloud classifier.
• Figure 2: Given an image coordinate $\bm{x}$, ARC finds the 4 nearest latent vectors, along with their relative position to $\bm{x}$. These are concatenated into a long descriptive vector which the decoder maps to an RGB colour. The reconstruction loss is computed between the ground-truth and predicted colour.
• Figure 3: Exp. 4: Image reconstruction quality when trained with various mini-batch ratios. Interestingly, both SIREN and ARC show strong fitting performance when only using a subset of the pixels at each optimization step. This allows for achieving a higher PSNR in less wall time.
• Figure 4: Feature locality exp.: Each column depicts a transformation on the latent vector positions, where the bottom row shows the resulting construction. No other changes are made to the ARC. The correspondence between the latent cloud transformation and the new reconstructions demonstrates how ARC encodes image features locally. In the rightmost two cases, two ARC, trained with a shared decoder, are mixed by masking specific latents or by simply stacking them.
• Figure 5: Image reconstruction results on the KODAK dataset. We compare ARC against SIREN and FINER, reporting PSNR as a function of model capacity. The vertical line represents the parameter count of a raw KODAK image, indicating the memory efficiency of different INR representations. SIREN and FINER experiments were bounded by computation costs, whereas ARC was not. ARC has a higher lower bound on memory than baselines but demonstrates superior scaling capabilities.