---
kind: concept
slug: hierarchical-vision-transformer
title: Hierarchical vision transformers
summary: >-
  Vision transformers that shrink the token grid in stages and widen the channels, as CNNs do. The variants differ in how they reduce the grid or the attention cost: PiT and P2T pool, EdgeViT and LITv2 mix local and global attention, HiFormer and SwinNet build on Swin pyramids. Trade-off: gains depend on recipe, hardware and task, and the numbers are author-reported.
tags: [vision-transformer, architecture, pooling, efficient-attention, mobile, image-classification, segmentation, salient-object-detection, medical-imaging, backbone]
papers: ["2103.16302", "2106.12011", "2205.03436", "2205.13213", "2207.08518", "2204.05585"]
related: [vision-transformer, efficient-vision-backbone, convolutional-neural-network, shifted-window-attention, local-global-attention, attention-mechanism, depthwise-separable-convolution, knowledge-distillation, semantic-segmentation, salient-object-detection, multimodal-fusion, multi-scale-feature-fusion]
status: draft
updated_at: 2026-11-01
---

## Overview

A plain Vision Transformer (ViT) keeps the same number of spatial tokens in
every layer [arxiv:2103.16302 §1]. A hierarchical variant instead reduces the
spatial grid in steps and increases the channel count, so later layers see a
smaller map with more features per position, as in a CNN [arxiv:2103.16302
§2] [arxiv:2103.16302 §3.2]. The shared idea is a sequence of stages, each
with its own spatial size, joined by a downsampling step. Whether the
downsampling is a pooling layer, a strided convolution or a patch merge is a
design choice that varies by variant. HiFormer, for example, uses patch merging
in its Swin branch [arxiv:2207.08518 §3.1.2].

Most later variants also change the attention itself, so that the cost falls
with the grid size. They do this by pooling the keys and values [arxiv:2106.12011
§III-B.1], by attending from a sparse set of delegate tokens [arxiv:2205.03436
§3.2], or by splitting the heads between local windows and pooled global
attention [arxiv:2205.13213 §4.1]. The sections below describe each one.

## Variants

### Pooling-based ViT (PiT)

PiT splits the transformer blocks into three stages with three spatial scales
[arxiv:2103.16302 §3.2]. Between stages, a pooling layer reshapes the spatial
tokens into a map, applies a depth-wise convolution that halves the spatial
size and widens the channels, and flattens the map back into tokens. The class
token has no position, so it passes through a linear layer that changes its
width instead [arxiv:2103.16302 §3.2]. Channel width per stage is set by the
head count of that stage [arxiv:2103.16302 §3.4] [arxiv:2103.16302 tab. 1].
Card: [Rethinking Spatial Dimensions of Vision Transformers](/papers/2103.16302).

Reported ImageNet-1k result at 224 x 224, same DeiT-style recipe: PiT-S
reaches 80.9% top-1 with 2.9 B FLOPs, against 79.8% and 4.6 B FLOPs for ViT-S
[arxiv:2103.16302 tab. 2] [arxiv:2103.16302 tab. 3]. Author-reported, single
runs, no variance given.

### Pyramid pooling attention (P2T)

P2T keeps global attention but shortens the keys and values. It average-pools
the token map at four ratios, adds a 3 x 3 depth-wise convolution to each
pooled map, flattens and concatenates the four maps, and uses the result as the
keys and values [arxiv:2106.12011 §III-B.1] [arxiv:2106.12011 eq. 2-4]. Its
four stages have feature strides 4, 8, 16 and 32 [arxiv:2106.12011 §III-A].
The feed-forward block is an inverted bottleneck with a depth-wise convolution
[arxiv:2106.12011 §III-B.2]. Card: [P2T: Pyramid Pooling Transformer for Scene
Understanding](/papers/2106.12011).

Reported results, all author-reported: at matched size P2T-Tiny reaches 79.8%
top-1 with 11.6 M parameters and 1.8 GFLOPs, against 78.7% and 13.1 M / 2.1
GFLOPs for PVTv2-B1. P2T-Small reaches 82.4%, against 82.0% for PVTv2-B2
[arxiv:2106.12011 tab. 2]. With Semantic FPN on ADE20K, P2T-Small reports 46.7
mIoU against 46.1 for PVTv2-B2 [arxiv:2106.12011 tab. 3]. With RetinaNet on
COCO, P2T-Small reports 44.4 AP against 43.8, and with Mask R-CNN box AP it is
45.5 against 45.3 [arxiv:2106.12011 tab. 4]. The gains over PVTv2 are therefore
between about 0.2 and 1.1 points in these tables. In the paper's 100-epoch
ablation, adding the depth-wise convolution to the pooled maps and the
inverted-bottleneck feed-forward raises top-1 from 74.7 to 79.5
[arxiv:2106.12011 tab. 9]. Those ablation numbers are not comparable to the
300-epoch main results.

### Local-global block per layer (EdgeViT)

Each EdgeViT layer is a local-global-local block [arxiv:2205.03436 §3.2]. First,
depth-wise and point-wise convolutions mix each token with its neighbours. Then
self-attention runs over one sampled delegate token per r x r window, so the
attention cost depends on the number of windows. Last, a transposed convolution
with kernel and stride r spreads each delegate's context back over its window.
The sample rates are 4, 2, 2 and 1 for the four stages [arxiv:2205.03436 §3.3].
The paper contrasts this with PVT and Twins, where every token issues a query
[arxiv:2205.03436 §3.2]. Card: [EdgeViTs: Competing Light-weight CNNs on Mobile
Devices with Vision Transformers](/papers/2205.03436).

The paper's target is on-device latency, measured on a phone CPU at FP32 with
batch 1 [arxiv:2205.03436 §4.1]. On that setup EdgeViT-XXS reports 74.4% top-1
at 32.8 ms, XS 77.5% at 54.1 ms and S 81.0% at 85.3 ms [arxiv:2205.03436
tab. 2]. MobileNet-v2 reports 72.0% at 33.3 ms, and EfficientNet-B0 77.1% at
52.1 ms. MobileNet-v3 0.75 is faster, at 23.0 ms, with 73.3% [arxiv:2205.03436
tab. 2]. So EdgeViT-XXS is more accurate than that CNN but slower on this phone.
In the ablation at about 0.5 GFLOPs, removing local aggregation drops top-1
from 74.4 to 72.7, and removing local propagation drops it to 73.9
[arxiv:2205.03436 tab. 4].

### Head-split local and global attention (LITv2, HiLo)

HiLo splits the heads of one attention layer into two groups [arxiv:2205.13213
§4.1] [arxiv:2205.13213 eq. 6]. The Lo-Fi heads, a fraction alpha of the total,
average-pool keys and values over s x s windows, while their queries still come
from every token. The Hi-Fi heads run standard attention inside non-overlapping
s x s windows, without shifting. The authors set alpha to 0.9 by grid search on
ImageNet-1K and use s = 2 for classification [arxiv:2205.13213 §5.1]. LITv2
places HiLo in stages 3 and 4 and removes the attention from stages 1 and 2
[arxiv:2205.13213 §4.3] [arxiv:2205.13213 app. A]. It replaces relative position
bias with a 3 x 3 depth-wise convolution inside each feed-forward block
[arxiv:2205.13213 §4.2]. Card: [Fast Vision Transformers with HiLo Attention](/papers/2205.13213).

Reported on an RTX 3090 with batch 64: LITv2-S reaches 82.0% top-1 at 1,471
images/s, against 81.3% at 961 images/s for Swin-Ti [arxiv:2205.13213 tab. 1].
On LITv2-S, plain multi-head self-attention reaches 82.3% at 1,293 images/s,
so HiLo buys speed for a small accuracy loss [arxiv:2205.13213 tab. 4]. On
ADE20K with Semantic FPN, HiLo reports 44.3 mIoU against 43.7 for plain
attention [arxiv:2205.13213 tab. 4]. The paper's own GPU table reports LITv2-S
at between 928 and 1,874 images/s across four GPUs, so absolute speed depends
on the card [arxiv:2205.13213 tab. 6].

### Swin-style pyramid with a CNN front end (HiFormer)

HiFormer is a 2D segmentation model with two pyramids side by side. A CNN with
three levels extracts local features, and a Swin Transformer with three levels
builds a multi-scale view on top of them [arxiv:2207.08518 §3.1.1]
[arxiv:2207.08518 §3.1.2]. Patch merging concatenates 2 x 2 neighbouring tokens,
doubling the width and halving the resolution [arxiv:2207.08518 §3.1.2]. The
largest and the smallest levels are then fused by a Double-Level Fusion (DLF)
module. Each outer level gets a class token, and that token attends to the other
level's patch tokens through cross-attention [arxiv:2207.08518 §3.1.3]. A
decoder turns the fused levels into a full-resolution label map
[arxiv:2207.08518 §3.2].

Three sizes, S, B and L, differ in the CNN depth and the DLF settings
[arxiv:2207.08518 tab. 1]. The reference B model reports 80.39 Dice and 14.70
HD on Synapse, against 79.13 Dice and 21.55 HD for Swin-Unet
[arxiv:2207.08518 tab. 2]. Removing DLF lowers Dice to 77.15 [arxiv:2207.08518
tab. 6]. Card: [HiFormer](/papers/2207.08518). All numbers are author-reported.

### Two-stream Swin-B pyramid for salient object detection (SwinNet)

SwinNet is a salient-object detector that needs a second, spatially paired
input: depth for RGB-D, or thermal for RGB-T [arxiv:2204.05585 §3.1]
[arxiv:2204.05585 §4.2]. Each input goes through its own Swin-B encoder, which
supplies the hierarchy [arxiv:2204.05585 §3.2]. At each level the two streams
are aligned with a shared spatial map and then re-weighted by channel attention
[arxiv:2204.05585 §3.3]. A separate edge branch built from the shallow depth
features guides a top-down decoder [arxiv:2204.05585 §3.4] [arxiv:2204.05585
§3.5]. Card: [SwinNet: Swin Transformer drives edge-aware RGB-D and RGB-T
salient object detection](/papers/2204.05585).

Reported RGB-D S-measure on NLPR is 0.941, against 0.931 for VST, and on NJU2K
it is 0.935, against 0.922 [arxiv:2204.05585 tab. 1]. In the backbone ablation
on NLPR, Swin-B gives 0.941, ResNet-50 gives 0.925 and ResNet-50 with a ViT-16
branch gives 0.932 [arxiv:2204.05585 tab. 3]. The model has 198.7 M parameters
and about 124.3 GFLOPs, and runs at about 10 FPS with I/O included
[arxiv:2204.05585 §4.4.5] [arxiv:2204.05585 tab. 7]. Its fusion and edge
modules are isolated in ablations [arxiv:2204.05585 tab. 4] [arxiv:2204.05585
tab. 5], so the hierarchy is an input to the method rather than the method
itself.

## When to use which

| Situation | Use | Why |
|---|---|---|
| You need a ViT-style classification backbone and FLOPs are fixed | PiT-style pooling, or P2T | At the same DeiT recipe PiT-S beats ViT-S with fewer FLOPs [arxiv:2103.16302 tab. 2]. P2T reports a similar gain over PVTv2 at matched size [arxiv:2106.12011 tab. 2] |
| You train with a weak recipe (vanilla, no CutMix) | PiT-style pooling | The gap is larger: PiT-S 73.3% vs ViT-S 68.7% [arxiv:2103.16302 tab. 2] |
| Dense prediction with a global-attention backbone at moderate cost | P2T | P2T-Small reports 46.7 mIoU on ADE20K against 46.1 for PVTv2-B2, at similar FLOPs [arxiv:2106.12011 tab. 3] |
| On-device classification with a latency budget on a phone CPU | EdgeViT-XXS or XS | EdgeViT-XXS is more accurate than MobileNet-v2 at similar latency [arxiv:2205.03436 tab. 2]. MobileNet-v3 0.75 is faster [arxiv:2205.03436 tab. 2], so pick EdgeViT only if you need the extra accuracy |
| GPU throughput with some global attention, and you can accept a small accuracy loss | HiLo (LITv2-S) | 1,471 images/s against 1,293 for plain attention, at 82.0 against 82.3 top-1 [arxiv:2205.13213 tab. 4] |
| You need the largest ImageNet accuracy and speed matters less | Plain attention, or check the numbers | Plain attention reaches 82.3 top-1 in the same LITv2-S setting [arxiv:2205.13213 tab. 4] |
| 2D medical segmentation with CNN-like local detail | HiFormer-B, or S for fewer parameters | On Synapse, B reports 80.39 Dice with 25.51 M parameters, against 79.13 Dice and 27.17 M for Swin-Unet [arxiv:2207.08518 tab. 2] [arxiv:2207.08518 tab. 5] |
| RGB-D or RGB-T salient object detection, with the second input present at test time | SwinNet | Reports the best RGB-D S-measure on NLPR and NJU2K among its listed baselines, and 0.941 against 0.931 for VST on NLPR [arxiv:2204.05585 §4.3] [arxiv:2204.05585 tab. 1]. Costs about 199 M parameters [arxiv:2204.05585 tab. 7] |
| Very small model budget | A CNN, or check the numbers | The paper says CNNs at the Ti scale beat both ViT-Ti and PiT-Ti [arxiv:2103.16302 §4.1] |
| Input larger than 224 x 224, or a non-square grid | Check first | HiFormer's experiments all use 224 x 224 inputs [arxiv:2207.08518 §4.2]. PiT's high-resolution result is a fine-tune with a cost penalty [arxiv:2103.16302 tab. 4]. HiLo's classification path assumes square grids that divide by s [arxiv:2205.13213 §4.1] |

## Implementation sketch

The common core is a loop over stages: run the transformer blocks at one
resolution, then downsample the spatial tokens and widen the channels. Where
variants differ is the downsampling step and how the class token is treated.
P2T and EdgeViT leave this loop unchanged and replace the attention inside each
block with a pooled or sampled version. HiLo keeps the loop and splits the heads
of each attention layer. HiFormer does not follow the loop exactly. It runs a
CNN pyramid and a Swin pyramid, and its fusion step replaces the final
downsample [arxiv:2207.08518 §3.1.3]. Its own pseudocode is on its card.

```pseudocode
# Hierarchical ViT forward pass, PiT-style [arxiv:2103.16302 §3.2]
# Input:  images (B, 3, 224, 224)
# Output: class logits (B, num_classes)

def hierarchical_vit(images):
    # Patch embedding gives a 27 x 27 grid in the paper's text [arxiv:2103.16302 §3.4]
    feature_map = patch_embed(images)
    # One learned class token, carried alongside the spatial tokens
    class_tokens = cls_token.expand(batch_size)
    for stage in range(num_stages):
        # Transformer blocks at this resolution [arxiv:2103.16302 tab. 1]
        feature_map, class_tokens = blocks[stage](feature_map, class_tokens)
        if stage < num_stages - 1:
            # Halve the grid and widen the channels between stages [arxiv:2103.16302 §3.2]
            feature_map, class_tokens = downsample[stage](feature_map, class_tokens)
    # Classify from the first class token [arxiv:2103.16302 §3.2]
    return head(layer_norm(class_tokens)[:, 0])
```

The downsample step is the variant-specific part. For PiT it is a depth-wise
convolution on the map and a linear layer on the class token. The paper does
not state the convolution's kernel size or padding [arxiv:2103.16302 §3.2].
For P2T the variant-specific part is the pooling pyramid inside attention, and
the paper does not state how pooled sizes are rounded when they do not divide
evenly [arxiv:2106.12011 §III-B.1].

## Contradictions

- **PiT-B patch size, within one paper.** The text gives patch 16 and stride 8
  for all PiT sizes [arxiv:2103.16302 §3.4]. Table 1 lists a 31 x 31 map for
  PiT-B, which those values do not produce at 224 input [arxiv:2103.16302
  tab. 1]. The reference code uses different values for B. Follow the table
  or the code, and check the choice before building B.
- **Robustness table label, within one paper.** The caption of the robustness
  table names PiT-B, but the PiT row is labelled PiT-S, and the text matches
  PiT-S [arxiv:2103.16302 tab. 6].
- **Where DLF output goes, within one paper.** Eq. 3 keeps the patch tokens of
  both levels unchanged, so the cross-level information is carried only in the
  class tokens [arxiv:2207.08518 eq. 3]. The decoder is said to consume the
  two patch grids [arxiv:2207.08518 §3.2]. The paper does not say how the
  class-token output reaches the decoder.
- **PVTv2-B1 detection baseline, across two papers.** P2T reports PVTv2-B1 at
  40.2 AP with RetinaNet on COCO [arxiv:2106.12011 tab. 4]. EdgeViT reports
  41.2 for the same backbone and the same parameter count of 23.8 M
  [arxiv:2205.03436 tab. 5]. The two cards do not explain the gap, and the
  PVTv2 paper was not checked. Do not cite either number as the PVTv2-B1
  baseline without that check.
- **PVTv2-B1 segmentation baseline, across two papers.** P2T reports 41.5 mIoU
  on ADE20K at 33.1 GFLOPs [arxiv:2106.12011 tab. 3]. EdgeViT reports 42.5 mIoU
  at 34.2 GFLOPs for the same backbone [arxiv:2205.03436 tab. 6]. Both use a
  512 x 512 input and a Semantic FPN head. The same caution applies.
- **SwinNet, claim against table.** The text says the RGB-T results are best on
  every metric [arxiv:2204.05585 §4.3.2]. Table 2 shows CGFNet with a higher
  F-measure on VT1000, 0.906 against 0.896 for SwinNet [arxiv:2204.05585
  tab. 2]. Use the table when ranking.

No other cross-paper disagreement is known. The variants report on different
tasks, ImageNet classification, COCO detection, ADE20K segmentation, medical
segmentation and salient object detection, so their numbers are not comparable
across tasks.

## Pitfalls

- **Gains depend on the recipe.** At the vanilla setting PiT-S leads ViT-S by
  4.6 points. With full DeiT plus distillation the lead is 0.7 points [arxiv:2103.16302 tab. 2].
- **Hidden cost at higher resolution.** PiT-B throughput falls to 82 img/s
  against 91 for ViT-B at 384 x 384, so the saving does not carry over to
  larger inputs [arxiv:2103.16302 tab. 4].
- **Under-specified downsampling.** Kernel size, padding and bias of the
  pooling convolution are not in the text [arxiv:2103.16302 §3.2]. Write the
  choice down explicitly when you implement it.
- **Detection comparisons.** The paper halves the input resolution for all
  backbones and does not state the exact size [arxiv:2103.16302 §4.2]. Its
  numbers are not directly comparable to standard COCO results.
- **Architecture search.** PiT was not found by neural architecture search, and
  the authors say it could improve with one [arxiv:2103.16302 §3.4].
- **Head count matters in the fusion branch.** HiFormer's ablation with 12 heads on
  the small level scores lower Dice than its default of 6 heads on both
  levels [arxiv:2207.08518 tab. 8]. The reported DLF gain is +3.24 Dice
  [arxiv:2207.08518 §5] [arxiv:2207.08518 tab. 6].
- **Two HiFormer rows match exactly.** The ResNet18 row of the backbone ablation
  and the no-DLF row of the DLF ablation both report 77.15 Dice and 16.88 HD
  [arxiv:2207.08518 tab. 6] [arxiv:2207.08518 tab. 7]. The paper does not
  explain this, so check it before citing either row.
- **Encoders dominate the cost.** HiFormer's cross-attention is linear in the
  token count, because only class tokens query. The level encoders before it
  are global self-attention over every token of a level [arxiv:2207.08518
  §3.1.3] [arxiv:2207.08518 tab. 1].
- **P2T text and table disagree on a detection gain.** The text claims P2T-Base
  leads Swin-S by 1.0 AP with RetinaNet [arxiv:2106.12011 §IV-C]. The table
  gives 46.1 against 44.5, a gap of 1.6 [arxiv:2106.12011 tab. 4].
- **P2T ablations are not the main recipe.** Its ablations train for 100 epochs
  and the baseline row has no pooling-convolution, no feed-forward convolution
  and no overlapping patch embedding [arxiv:2106.12011 tab. 9]. Use the
  300-epoch table rows for comparison with other models.
- **Pooling type matters.** In P2T's ablation, average pooling gives 74.7 top-1
  against 73.0 for max pooling and 73.8 for depth-wise convolution as the pooling
  operation [arxiv:2106.12011 tab. 7].
- **Energy metric favours small models.** EdgeViT's efficiency value is top-1
  divided by energy, which the paper notes favours less accurate models. That is
  why it compares only Pareto-optimal points [arxiv:2205.03436 §4.1]
  [arxiv:2205.03436 tab. 3].
- **Latency and energy come from different phones.** Latency is measured on a
  Galaxy S21 and energy on a Snapdragon HDK [arxiv:2205.03436 §4.1]. Do not
  derive energy from latency and power.
- **Figure text and table disagree.** The EdgeViT Figure 1 caption gives 33.3 ms
  for the lightest model, while its table gives 32.8 ms for EdgeViT-XXS and 33.3
  ms for MobileNet-v2 [arxiv:2205.03436 fig. 1] [arxiv:2205.03436 tab. 2].
- **EdgeViT detection gain in the text.** The text gives +1.2 AP^m for EdgeViT-S
  over PVTv2-B1 on Mask R-CNN [arxiv:2205.03436 §4.2.1]. The table gives 41.0
  against 38.8, a gap of 2.2 [arxiv:2205.03436 tab. 5].
- **EdgeViT ablation rescales the model.** Ablation variants are rescaled to
  about 0.5 GFLOPs by changing layer counts, so they are not the released models
  [arxiv:2205.03436 §4.1.1] [arxiv:2205.03436 app. A.2].
- **Plain attention is the accuracy reference for HiLo.** It scores 0.3 top-1 more
  on ImageNet, and the paper's gain is speed and FLOPs [arxiv:2205.13213 tab. 4].
- **Window size and resolution in HiLo.** The paper says a window of 4 speeds up
  high-resolution inputs with a small accuracy loss [arxiv:2205.13213 §5.2]
  [arxiv:2205.13213 app. C]. In its detection table, LITv2-S with window 4 drops
  RetinaNet AP from 44.0 to 43.7 [arxiv:2205.13213 tab. 2].
- **Lo-Fi head count is rounded down.** Alpha times the head count is not an
  integer in general. The paper does not say how to round it [arxiv:2205.13213
  §4.1].
- **HiLo recipe is partly missing.** The paper gives warm-up, clipping, label
  smoothing and mixup only as "the same as in LITv1" [arxiv:2205.13213 §5.1].
- **SwinNet feature levels.** The paper names four levels but does not say which
  Swin outputs they are [arxiv:2204.05585 §3.2]. An implementation that takes
  the four stage outputs may not match the released model, so check this choice
  against the card.
- **SwinNet checkpoint choice.** The paper does not say how the reported
  checkpoint was picked [arxiv:2204.05585 §4.2]. Ask which validation split was
  used before you trust a reported number.
- **SwinNet speed.** About 10 FPS is reported with I/O and preprocessing included,
  on hardware not stated for that figure [arxiv:2204.05585 §4.4.5]. Training
  took about 26 hours on one RTX 2080 Ti [arxiv:2204.05585 §4.2].
- **Hardware-specific speed.** Throughput depends on the GPU or CPU. The
  throughput rankings in each paper hold only for the setup it names
  [arxiv:2205.13213 §5.1] [arxiv:2205.03436 §4.1] [arxiv:2106.12011 tab. 2].
- **Input size.** P2T and EdgeViT classification use 224 x 224 inputs
  [arxiv:2106.12011 §IV-A] [arxiv:2205.03436 §4.1]. Check the input size of
  each detection and segmentation row before comparing it with another.

## Papers

- [Rethinking Spatial Dimensions of Vision Transformers](/papers/2103.16302): pooling between transformer stages, shrinking the token grid while widening channels; reports gains over ViT at equal FLOPs on ImageNet, detection and robustness.
- [P2T: Pyramid Pooling Transformer for Scene Understanding](/papers/2106.12011): multi-scale average pooling of keys and values, with depth-wise convolutions on each pooled map; reports ImageNet, ADE20K and COCO results against PVTv2 and Swin.
- [EdgeViTs: Competing Light-weight CNNs on Mobile Devices with Vision Transformers](/papers/2205.03436): a local-global-local block with delegate-token attention; reports phone-CPU latency and energy against light CNNs and transformers.
- [Fast Vision Transformers with HiLo Attention](/papers/2205.13213): splits heads into high-frequency window attention and low-frequency pooled attention inside a four-stage backbone; reports ImageNet, COCO and ADE20K results against Swin and other efficient attention.
- [HiFormer: Hierarchical Multi-scale Representations Using Transformers for Medical Image Segmentation](/papers/2207.08518): a CNN pyramid and a Swin pyramid fused by class-token cross-attention; reports Synapse, skin-lesion and SegPC results, with an official repository that differs from the paper in several places.
- [SwinNet: Swin Transformer drives edge-aware RGB-D and RGB-T salient object detection](/papers/2204.05585): two Swin-B encoders, per-level alignment and channel re-calibration, and an edge-guided decoder; reports RGB-D and RGB-T results against other salient object detectors.