# CDeC-Net: Composite Deformable Cascade Network for Table Detection > CDeC-Net combines cascade Mask R-CNN with a dual composite backbone and deformable convolutions for table detection, reporting results at or near prior published figures across seven public benchmarks with a single trained model. - Authors: Hunter Heidenreich - Published: 2020-08-25 - Updated: 2026-03-24 - Paper: Preprint (arXiv) - Paper: Published (ICPR 2020) - Code: CDeCNet Code - Source: Bevaya Labs, --- ## TL;DR CDeC-Net extends Cascade Mask R-CNN with a composite dual-backbone architecture and deformable convolutions to improve table detection accuracy across a range of document types and IoU thresholds. A single model trained on IIIT-AR-13K produces results at or near dataset-specific baselines across seven benchmark datasets without dataset-specific tuning. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Method}}$: The headline contribution is a new detection architecture. The paper proposes CDeC-Net, combining three existing components (cascade detection, composite backbones, deformable convolution) in a new configuration tailored for table detection. The bulk of the paper consists of ablations and baseline comparisons that validate design choices. **Secondary:** $\Psi\_{\text{Evaluation}}$: A notable secondary contribution is the breadth of evaluation: seven benchmark datasets are tested under multiple IoU thresholds and diverse experimental protocols, including a single-model evaluation variant. ## What is the motivation? Table detection in document images is complicated by high intra-class variability (tables differ widely in layout and ruling style) and inter-class similarity (figures, graphs, and equations can visually resemble tables). At the time of this work, existing deep learning approaches suffered from three interconnected problems: 1. **Backbone mismatch:** Standard backbones pre-trained for image classification (ImageNet) are not necessarily well suited to extracting features for document table detection. Training a deeper, higher-capacity backbone from scratch is expensive. 2. **Fixed geometric structure:** Standard convolutions sample from a fixed grid, limiting the model's ability to handle the geometric variation typical of tables (spanning cells, rotated content, irregular borders). 3. **IoU threshold sensitivity:** Most detectors train with an IoU threshold of 0.5 for positive sample assignment. At higher evaluation thresholds the number of valid positives drops sharply, leading to degraded performance and imprecise bounding boxes. Additionally, prior methods typically trained and evaluated a separate model per benchmark. The authors argue for a generic solution: a single model that works across diverse datasets. ## What is the novelty? CDeC-Net composes three independently proposed techniques into a unified architecture: **Cascade Mask R-CNN:** A multi-stage extension of Faster R-CNN in which each successive stage is trained at a higher IoU threshold (0.5, 0.6, 0.7). Each stage regresses bounding boxes that serve as proposals for the next stage. This progressive refinement maintains enough positive samples at higher thresholds while producing tighter boxes. **Composite (dual) backbone:** Inspired by CBNet, two ResNeXt-101 backbones are stacked in parallel. The "assistant" backbone feeds its stage-$l$ output into the corresponding stage of the "lead" backbone via a composite connection $g(\cdot)$: $$x\_{bl}^l = F\_{bl}^l\!\left(x\_{bl}^{l-1} + g\!\left(x\_{ba}^l\right)\right), \quad l \ge 2$$ This provides the lead backbone with richer features derived from a second forward pass without retraining a substantially larger single network. **Deformable convolution:** Standard convolution samples over a fixed regular grid $R$: $$y(p\_0) = \sum\_{p\_n \in R} w(p\_n) \cdot x(p\_0 + p\_n)$$ Deformable convolution augments each sampling location with a learned offset $\Delta p\_n$: $$y(p\_0) = \sum\_{p\_n \in R} w(p\_n) \cdot x(p\_0 + p\_n + \Delta p\_n)$$ The offsets are themselves learned end-to-end, allowing each neuron to adapt its receptive field to the shape and scale of the object being detected. Both dual backbones use deformable convolution throughout. The combination of these three components is applied to all benchmark datasets under a unified training regime, including a single-model variant (CDeC-Net$^\ddagger$) trained only on IIIT-AR-13K and optionally fine-tuned on each target dataset. ## What experiments were performed? **Datasets:** ICDAR-2013 (238 test images), ICDAR-POD-2017 (817 test images, multi-class), UNLV (424 test images), Marmot (1K test images, English and Chinese), ICDAR-2019 cTDaR (439 test images), TableBank (LaTeX and Word splits, 1K test each), PubLayNet (11K test images, multi-class). **Metrics:** Recall (R), Precision (P), F1 score, and mean Average Precision (mAP) at IoU thresholds matching each benchmark's established protocol. **Baselines:** Per-benchmark state-of-the-art methods at time of submission, including DeCNT (deformable Faster R-CNN), GOD (Mask R-CNN variant), YOLO-based detectors, and Mask/Faster R-CNN variants from the PubLayNet and TableBank papers. **Implementation:** MMDetection (PyTorch), ResNeXt-101 backbone pre-trained on MS-COCO with FPN, input resolution 1200$\times$800, initial learning rate 0.00125 with warmup (0.0033 for 500 iterations), decay at epochs 25 and 40, 50 total epochs (8 epochs for large datasets such as PubLayNet and TableBank). Batch size 1 on a single NVIDIA GeForce RTX 2080 Ti (12 GB). Multi-scale testing at 7 scales; a detection is kept only if it appears in at least 4 of the 7 scale runs. **Ablation (Table IX, trained on Marmot, tested on ICDAR-2013 at IoU 0.5):** | Model | F1 | |---|---| | Cascade Mask R-CNN + ResNeXt-101 | 0.981 | | + Composite backbone | 0.984 | | + Deformable convolution (CDeC-Net) | 1.000 | **IoU robustness (Table VIII/XIII, single best model per dataset):** CDeC-Net maintains high F1 from 0.5 to 0.8 on ICDAR-2013 and ICDAR-2019. Performance on UNLV and ICDAR-2013 drops noticeably at IoU 0.9 (F1 0.557 and 0.660 respectively), while ICDAR-2019 retains F1 0.915 at IoU 0.9. ## What are the outcomes/conclusions? **Dataset-specific models (Table III):** CDeC-Net exceeds prior state of the art on five of seven benchmarks and comes within 0.1% F1 on a sixth (ICDAR-2019): | Dataset | F1 | mAP | |---|---|---| | ICDAR-2013 | 1.000 | 1.000 | | ICDAR-2017 | 0.947 | 0.912 | | ICDAR-2019 | 0.944 | 0.922 | | UNLV | 0.938 | 0.912 | | Marmot | 0.952 | 0.911 | | TableBank | 0.987 | 0.976 | | PubLayNet | 0.978 | 0.967 | ICDAR-2017 is the one dataset where CDeC-Net falls 2.4% F1 below the YOLO-based state of the art (0.971), which the authors attribute to those methods being specifically optimized for ICDAR-2017. **Single model (CDeC-Net$^\ddagger$):** The IIIT-AR-13K-trained model reaches F1 figures close to or above dataset-specific baselines on all datasets after optional fine-tuning, notably exceeding those baselines on ICDAR-2019 and PubLayNet (Table I). **Failure modes:** The model merges closely spaced tables into a single bounding box and occasionally produces false positives on graph-like structures that resemble tables. No error bars or statistical significance tests are reported. **Limitations acknowledged:** The ICDAR-2017 gap is noted. The authors do not discuss computational overhead relative to simpler baselines or the practical cost of multi-scale inference at 7 scales. ## Reproducibility ### Models - **Architecture:** Cascade Mask R-CNN with dual composite ResNeXt-101 backbones and deformable convolution, plus FPN. Three cascade stages at IoU thresholds 0.5, 0.6, and 0.7. - **Backbone:** ResNeXt-101 (blocks 3-4-23-3 configuration), pre-trained on MS-COCO with FPN. Two identical instances connected via composite connections. - **Code and weights:** The paper states code and models will be publicly released. The repository has moved to https://github.com/madhav1ag/CDeCNet and is released under the MIT license. Pretrained model weights are included in the repository. ### Algorithms - **Framework:** MMDetection (PyTorch) - **Optimizer:** SGD (implied by MMDetection defaults; not explicitly stated) - **Learning rate:** 0.00125 initial; decay at epochs 25 and 40; warmup factor 0.0033 for first 500 iterations - **Epochs:** 50 for small/medium datasets; 8 for PubLayNet and TableBank; 12 for fine-tuning - **Batch size:** 1 - **Input resolution:** 1200$\times$800 (aspect ratio preserved) - **Anchor ratios:** 0.5, 1.0, 2.0; single anchor scale of 8 - **Test-time augmentation:** Multi-scale at 7 scales (3 smaller, original, 3 larger); consensus voting requiring presence in at least 4 of 7 scales ### Data - **Primary training set for single model:** IIIT-AR-13K (9K training, 2K validation, 2K test; 5 classes including table, figure, natural image, logo, signature) - **Benchmark datasets used:** ICDAR-2013, ICDAR-POD-2017, UNLV, Marmot, ICDAR-2019 (cTDaR), TableBank (LaTeX, Word, both splits), PubLayNet (all public) - **Annotation quality:** Ground truth for TableBank and PubLayNet is automatically generated; annotations for the other benchmarks are human-curated ### Evaluation - **Metrics:** Recall, Precision, F1, and mAP; IoU threshold varies per benchmark protocol (0.5 standard; some benchmarks use 0.6, 0.8, or 0.9) - **Protocol:** The authors replicate the exact training/fine-tuning/test splits and IoU thresholds of each compared paper to ensure fair comparison - **Statistical rigor:** No error bars, confidence intervals, or multiple-run reporting. Results are from single training runs per configuration. - **Limitations not acknowledged:** Multi-scale test-time augmentation at 7 scales substantially increases inference cost; this is not discussed relative to baselines that do not use TTA. ### Hardware - **Training GPU:** Single NVIDIA GeForce RTX 2080 Ti (12 GB VRAM) - **GPU-hours:** Not reported - **Inference:** Multi-scale testing at 7 scales implies non-trivial inference overhead; latency not reported - **Deployment:** The 12 GB training requirement suggests a mid-range workstation GPU is sufficient for training; inference requirements are not specified ## BibTeX ```bibtex @inproceedings{agarwal2020cdecnet, title = {{CDeC-Net}: Composite Deformable Cascade Network for Table Detection in Document Images}, author = {Agarwal, Madhav and Mondal, Ajoy and Jawahar, C. V.}, booktitle = {2020 25th International Conference on Pattern Recognition (ICPR)}, year = {2021}, pages = {9491--9498}, doi = {10.1109/ICPR48806.2021.9411922} } ```