UniFormer-S COCO top-down pose - acaua mirror (pure-PyTorch port)
==================================================================

This product includes:

1. PORTED SOURCE CODE: a pure-PyTorch port of the UniFormer-pose
   architecture (located in the acaua repository at
   src/acaua/adapters/uniformer/pose/) is a derivative work of:

     - Sense-X/UniFormer (Apache-2.0)
       https://github.com/Sense-X/UniFormer @ main
       Files derived from:
         pose_estimation/mmpose/models/backbones/uniformer.py
           (the global-attention UniFormer backbone, parameter-tree
            identical to the detection variant — we reuse acaua's
            UniFormer2DDense with hybrid=False, windows=False)
         pose_estimation/mmpose/models/keypoint_heads/top_down_simple_head.py
           (TopDownSimpleHead: 3x ConvTranspose2d-stride-2 + BN + ReLU
            + 1x1 final conv)
         pose_estimation/mmpose/core/evaluation/top_down_eval.py
           (keypoints_from_heatmaps with post_process='default' branch:
            argmax + sign(neighbor_diff) * 0.25 shift)

2. CONVERTED WEIGHTS: the model.safetensors file in this mirror is a
   key-remapped conversion of the upstream pretrained checkpoint:

     - upstream repo:      Sense-X/UniFormer (Apache-2.0)
     - upstream weights:   Google Drive file id
                           162R0JuTpf3gpLe1IK6oxRoQK7JSj4ylx
     - upstream filename:  top_down_256x192_global_small.pth
     - upstream SHA256:    d77059e3e9322c0e20dc89dc0cf2a583ffe2ced7d3e9b350233738add570bc30
     - upstream paper:     Li et al., "UniFormer: Unifying Convolution
                           and Self-attention for Visual Recognition",
                           arXiv:2201.09450 (ICLR 2022)

   Key remap rule (eager, at conversion time):
       backbone.*       ->  backbone.*       (no change)
       keypoint_head.*  ->  head.*

   Conversion was performed by scripts/convert_uniformer_pose.py in the
   acaua repository. The adapter loads the safetensors file under
   load_state_dict(strict=True) — key-naming drift between upstream and
   the port is a hard error at load time.

3. BUNDLED DETECTOR: this mirror's adapter loads
   CondadosAI/rtmdet_t_coco at adapter-init time so callers get the
   full top-down pipeline (detect persons -> per-person crop -> pose ->
   inverse-warp) from a single predict(image) call. RTMDet-tiny carries
   its own attribution chain — see CondadosAI/rtmdet_t_coco's NOTICE
   file.

Mirrored on 2026-04-25 by CondadosAI.

License
-------
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.