cstr/tps-loc-GGUF

TPS Localization Network (GGUF)

Thin-Plate Spline localization CNN for document dewarping. Predicts 20 control point coordinates from a document image, which are then used to compute a TPS warp that straightens curved/distorted text.

Architecture

PaddleOCR RARE "small" variant (~108K params):

• Conv0: 3->16, 3x3 + BN(folded) + ReLU + MaxPool2x2
• Conv1: 16->32, 3x3 + BN(folded) + ReLU + MaxPool2x2
• Conv2: 32->64, 3x3 + BN(folded) + ReLU + MaxPool2x2
• Conv3: 64->128, 3x3 + BN(folded) + ReLU + AdaptiveAvgPool(1)
• FC1: 128->64 + ReLU
• FC2: 64->40 (20 control points x 2 coords)

Files

File	Size	Description
tps-loc-f32.gguf	424 KB	F32 weights, 108K params

Source

Extracted from PaddleOCR rec_mv3_tps_bilstm_att_v2.0 recognition model (Apache-2.0). BatchNorm folded into conv weights at conversion time.

Usage

Parity

C++ vs Python reference: cos=1.000000, max_abs=0.000000 (exact F32 match).

Provenance and EU AI Act Art. 53 note

• Upstream model: PaddlePaddle/PaddleOCR.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.