Feature Map Convergence Evaluation for Functional Module

TL;DR

This paper tackles the lack of independent evaluation for the functional modules within autonomous driving perception models by introducing the Feature Map Convergence Score (FMCS) and the Convergence Quantification Indicator (CQI). It defines a principled workflow: quantify model convergence with CQI, segment training into $K$ convergence phases, derive backbones’ feature maps at epoch markers, generate the FMCS-Dataset, and train FMCE-Net to predict FMCS from these feature maps. The authors demonstrate high predictive accuracy of FMCE-Net across multiple datasets and backbones, and visualize Grad-CAM heatmaps to show how convergence corresponds to more localized and meaningful feature focus. This framework offers a new, quantitative paradigm for assessing training maturity of modular components in perception DNNs and could guide targeted optimization in autonomous driving systems. The work lays groundwork for independent module evaluation and points to future extensions to more complex, serially connected architectures.

Abstract

Autonomous driving perception models are typically composed of multiple functional modules that interact through complex relationships to accomplish environment understanding. However, perception models are predominantly optimized as a black box through end-to-end training, lacking independent evaluation of functional modules, which poses difficulties for interpretability and optimization. Pioneering in the issue, we propose an evaluation method based on feature map analysis to gauge the convergence of model, thereby assessing functional modules' training maturity. We construct a quantitative metric named as the Feature Map Convergence Score (FMCS) and develop Feature Map Convergence Evaluation Network (FMCE-Net) to measure and predict the convergence degree of models respectively. FMCE-Net achieves remarkable predictive accuracy for FMCS across multiple image classification experiments, validating the efficacy and robustness of the introduced approach. To the best of our knowledge, this is the first independent evaluation method for functional modules, offering a new paradigm for the training assessment towards perception models.

Figures (6)

• Figure 1: Feature Map Convergence Evaluation Framework. This framework is designed to generate a standardized Feature Map Convergence Score (FMCS) for the output feature maps of functional modules, such as the backbone, based on the analysis of the Loss Sequence from the task head. The Feature Map Evaluation Network (FMCE-Net) is trained on the Feature Map Convergence Score Dataset, serving as a tool to assess the maturity of the training of functional modules.
• Figure 2: CQI and FMCS Generation Flowchart. The CQI serves as a quantitative measure to assess the fluctuation extent of the Loss Sequence. Throughout the training process, K epoch markers are identified that represent the convergence phases that are evenly divided, with FMCS 1, …, $K$ for feature maps corresponding to $E_1$, …, $E_K$.
• Figure 3: In-Depth Visualization of Training Loss Curves and Epoch Markers. This figure shows the analysis of Convergence Metrics during ResNet-50 Training on the Mini-ImageNet Dataset over 150 Epochs. (a) presents the original and smoothed Loss Curve observed in a detailed view of the fluctuations around local region of convergence. (b) illustrates the trend of CQI around convergence threshold. (c) demonstrates a schematic representation epoch markers to signify convergence phases. (d) shows the uniform division of 10 convergence phases on logarithmically smoothed Loss Curve.
• Figure 4: Grad-CAM Heatmaps Across 10 Convergence phases of MNIST. Visualizations of feature maps, derived from ResNet-50 (R) and ShuffleNet v2 (S) backbones, same for the two images below. As the FMCS values escalate, the regions of high contribution increasingly localize, indicative of improved feature extraction efficiency.
• Figure 5: Grad-CAM Heatmaps Across 10 Convergence phases of CIFAR-10. Evidently, ResNet-50 does not exhibit a concentrated focus, and ShuffleNet erroneously focuses on the background when extracting features from images labeled with 'dog'. These lead to a relatively lower FMCS predictive accuracy.