Alias-Free Along the Ring: Exact Yaw Equivariance for 360° Panoramic Detection

A 360° camera that turns about its vertical axis produces a new panorama that is an exact horizontal wrap-around shift of the old one. No new content enters the frame and none leaves. In equirectangular projection (the standard "world map" layout of a panorama, with longitude across and latitude down), the left and right edges are physically the same place. This is the one common real-world setting where translation behaves the way the theory of alias-free networks assumes. Yet panoramic detectors use ordinary ReLU networks with unfiltered downsampling. They wrap the edges ("circular padding") only to avoid splitting objects at the seam. We propose to make the backbone and heatmap head of a panoramic detector exactly equivariant to arbitrary, including sub-pixel, yaw rotations. Equivariant here means that rotating the camera shifts the outputs by the same amount and otherwise leaves them unchanged.

The key design choice is to remove aliasing only along the periodic horizontal axis. Aliasing is the folding of high frequencies into false low frequencies, and ReLU-like activations create it by generating new high frequencies. The vertical axis is bounded and distorted, so it is left free to use absolute position. This one-dimensional treatment is what the physics of the sensor requires. It also roughly halves the extra cost of current alias-free networks, because each activation needs extra resolution in one direction rather than two. We would then measure, for the first time, how much a panoramic detector's accuracy swings under small camera rotations, and whether eliminating activation-induced aliasing closes the gap that padding fixes and blurred pooling leave open.

Core Claim

Alias-Free Convnets makes a classifier exactly invariant to integer and fractional shifts. It does this by replacing ReLU with low-degree polynomials, whose bandwidth growth is known, filtering after each activation, and treating the image as a torus with circular convolutions. It reports 100% shift consistency on ImageNet, but only for circular shifts. The authors acknowledge that this guarantee degrades into an approximation for real "crop-shifts," where new pixels enter at the border. The torus is an artificial assumption for planar photos.

Our claim is that, for equirectangular panoramas and LiDAR range images, the cylinder topology is real along azimuth. The certificate therefore becomes physically meaningful there, and only along that axis. A network that is alias-free and circularly padded horizontally, but conventional vertically, is exactly equivariant to camera yaw. It also keeps the vertical position cues that structured scenes need: floor below, ceiling above.

A further observation removes an apparent obstacle raised by the latitude-distortion problem. Equirectangular rows near the poles are oversampled horizontally, but a yaw rotation is the same angular shift in every row. A row-wise filter with a single cutoff therefore preserves exact yaw equivariance at every latitude. Latitude-dependent filtering may matter for accuracy, but it is not needed for the guarantee. We test this rather than assume it.

The closest work differs in identifiable ways:

  • Alias-Free Convnets and Alias-Free ViT stay in planar classification, filter in both dimensions, and pay large compute costs. The ViT version reports 487 versus 69 training hours.
  • StyleGAN3 and Alias-Free LDM supply the upsample–activate–downsample method we reuse. Both are generative and planar.
  • Panoramic detectors use ring padding with standard activations and never measure shift stability. These include Eliminating the Blind Spot, the R-CenterNet detector in PANDORA, and the FPN baselines in 360-Indoor.

Research Program

We build an "azimuth-alias-free" backbone in four parts.

First, the backbone. Every convolution uses circular padding horizontally and zero padding vertically. Every pointwise nonlinearity is wrapped in a horizontal-only sequence: upsample along azimuth, activate, apply a low-pass filter along azimuth, downsample. Every stride-2 stage applies an FFT-based ideal low-pass filter along azimuth before subsampling. Vertical downsampling is conventional. We test two activation choices. One is degree-2 polynomials, following Alias-Free Convnets: with upsampling by 2, a degree-2 polynomial's output fits exactly. The other is a filtered GELU, following Alias-Free ViT. It is only approximately alias-free, but it avoids the known instability of polynomial training.

Second, normalization. Normalization uses statistics pooled across azimuth within each row, or across the whole layer, so that per-pixel normalization does not reintroduce aliasing.

Third, vertical position. Elevation is supplied explicitly through CoordConv-style elevation channels. This makes concrete the separation that LiDARGen adopted informally, where it combined circular convolutions with coordinate channels.

Fourth, the detection head. The head is a dense, CenterNet-style heatmap and regression head built from the same operators. Its outputs are therefore yaw-equivariant before peak-finding. Shift Equivariance in Object Detection found heatmap detectors to be the most shift-fragile (ΔAP50 of 8.1 at 1 pixel on COCO). We report equivariance at two points: on the heatmaps before peak-finding, and on final AP after non-maximum suppression. This separates what the architecture fixes from what discrete peak selection breaks.

All models are pretrained on ImageNet-1K with the same recipe. For pretraining, planar images are treated as periodic horizontally. We pretrain because 360-Indoor found that pretrained standard models beat specialized spherical networks trained from scratch. We then fine-tune for detection.

Generating truly fractional yaw rotations at test time requires care. FFT-resampling a real panorama would assume the property we are testing. Instead, we start from the high-resolution panorama, apply an integer circular shift at that resolution, and downsample with a high-quality filter to network resolution. This yields genuine sub-pixel rotations of the scene.

Baselines

The naive reference is a standard ImageNet-pretrained detector with zero padding. Two strong published systems anchor accuracy. On 360-Indoor, FPN with a ResNet-101 backbone reports 33.6% mAP (360-Indoor). On PANDORA, R-CenterNet with Panoramic Rotation Data Augmentation (PRDA) reports 22.7 AP50, up from 19.9 for Sph-CenterNet (PANDORA). For neither dataset did we find a reported ΔAP under yaw rotation on Lacuna. These baselines fix the accuracy we must not lose. They do not tell us anything about stability.

The strongest competitor on the actual question is our reimplementation of R-CenterNet with three cheap fixes combined: circular padding, BlurPool before every downsampling (Making Convolutional Networks Shift-Invariant Again), and PRDA during training. It removes the seam and the downsampling aliasing that prior work already knew how to remove. Only activation-induced aliasing remains, so it isolates our contribution. Every model uses the same backbone width and depth, the same pretraining data and epochs, the same detection fine-tuning schedule, and the same hyperparameter search budget. We report wall-clock time alongside accuracy so that the comparison is honest about compute.

The proposal beats this baseline if it meets three conditions together. Its ΔAP50 under fractional yaw rotations must be at least halved relative to the baseline. Its heatmap consistency (shift-PSNR, as used in Alias-Free LDM) must approach numerical precision under both integer and fractional shifts. Its mean AP must stay within one point.

We also compare against the full two-dimensional alias-free design, adapted from Alias-Free Convnets, with circular padding vertically as well. That paper reports 5× slower training and 6.5× slower inference on classification. We expect the horizontal-only version to match its yaw equivariance at clearly lower cost. We also expect it to match or exceed its accuracy, since wrapping the vertical axis connects the top of the panorama to the bottom, which is physically wrong.

Two ablations isolate the key ingredient. The first removes only the activation filtering while keeping horizontal FFT downsampling and circular padding; this isolates activation aliasing. The second removes the elevation channels; this tests whether vertical position cues are needed once the horizontal anchors are gone.

Why not just…?

Why not just train with yaw rotation augmentation? Why do deep convolutional networks generalize so poorly to small image transformations? shows that augmentation does not produce invariance to translations not seen in training. In PANDORA, PRDA together with custom heatmaps added only 2.8 AP50, with stability unmeasured. Augmentation is still included in our strongest baseline. If that baseline already shows near-zero fractional ΔAP, our premise is wrong for this domain and we will report that.

Why not just add circular padding and BlurPool? This is the real threat. On a true cylinder, circular padding fully removes the boundary cue that On Translation Invariance in CNNs and Padding Aware Neurons identify. The open question is whether the remaining activation aliasing matters at detection scale. Shift Equivariance in Object Detection found that BlurPool removed only about 30% of the variance in planar detectors, but that measurement mixed in padding effects. Our first ablation separates them cleanly. If the activation filter adds little beyond this baseline, the honest conclusion is that panoramas need only cheap fixes. We would then publish that result together with the measurement protocol.

Why not just average predictions over several rotated copies at test time? Averaging over N rotations costs N forward passes. It also covers only the sampled rotations, so stability is still not guaranteed at the rotations in between. Shift Equivariance in Object Detection reports that this does not remove the underlying lack of equivariance. We include test-time averaging with N matched to our method's inference cost. This shows whether spending the same compute on averaging buys the same stability.

Proposed Experiments

The main experiments run on 360-Indoor (3,335 panoramas, 37 classes) and PANDORA (3,000 panoramas, 47 classes, rotated field-of-view boxes). Boxes that cross the seam are re-wrapped after each rotation. Yaw rotations are sampled at quarter-pixel steps within one network pixel, and at larger offsets up to a full revolution.

We report mean AP and AP50; ΔAP50, the spread between maximum and minimum AP across rotations; shift-PSNR of backbone features and heatmaps; the rate of class swaps for matched objects; and throughput.

As a secondary check that the guarantee is not specific to camera images, we evaluate feature-level yaw consistency of range-image encoders on KITTI-360 LiDAR scans. Rotating the LiDAR sensor is likewise an exact azimuthal shift of its range image.

A last diagnostic tests the latitude argument. We compare a filter with one cutoff for all rows against a latitude-dependent cutoff. The prediction is that both are equally equivariant and differ only in accuracy near the poles.

Three results are decisive. The first is exactness: whether the proposal achieves near-perfect fractional yaw consistency where the circular-plus-BlurPool baseline does not. The second is whether that gap appears in ΔAP after peak-finding and non-maximum suppression, or vanishes at that stage. The third is whether horizontal-only filtering retains accuracy at substantially lower cost than full two-dimensional alias-free filtering.