PaperlayerView exact revision on arXiv

Source-linked AI summary

SlimYOLOv3: Narrower, Faster and Better for Real-Time UAV Applications

Pengyi Zhang, Yunxin Zhong, Xiaoqiong Li

arXiv:1907.11093v1cs.CV

TL;DR

UAV object detection must operate under limited memory and computing power. The paper learns slimmer YOLOv3 detectors through sparsity-guided channel pruning and reports large efficiency gains with comparable accuracy on VisDrone2018-Det.

  • Problem

    Real-time UAV object detection is challenging because embedded platforms have limited memory and computing power.

  • Method

    The method imposes L1 regularization on channel scaling factors, prunes less informative convolutional channels, and fine-tunes the resulting SlimYOLOv3 detector.

  • Results

    SlimYOLOv3 achieves comparable detection accuracy to YOLOv3 while reducing FLOPs by ~90.8%, reducing parameter size by ~92.0%, and running ~2 times faster.

  • Takeaways & Limitations

    SlimYOLOv3 is presented as more suitable for real-time object detection on UAVs because it is narrower, faster, and more efficient than YOLOv3.

  • Takeaways & Limitations

    VisDrone2018-Det has substantial category imbalance that is not managed in the experiments, contributing to lower overall performance.

Abstract

from arXiv · show

Drones or general Unmanned Aerial Vehicles (UAVs), endowed with computer vision function by on-board cameras and embedded systems, have become popular in a wide range of applications. However, real-time scene parsing through object detection running on a UAV platform is very challenging, due to limited memory and computing power of embedded devices. To deal with these challenges, in this paper we propose to learn efficient deep object detectors through channel pruning of convolutional layers. To this end, we enforce channel-level sparsity of convolutional layers by imposing L1 regularization on channel scaling factors and prune less informative feature channels to obtain "slim" object detectors. Based on such approach, we present SlimYOLOv3 with fewer trainable parameters and floating point operations (FLOPs) in comparison of original YOLOv3 (Joseph Redmon et al., 2018) as a promising solution for real-time object detection on UAVs. We evaluate SlimYOLOv3 on VisDrone2018-Det benchmark dataset; compelling results are achieved by SlimYOLOv3 in comparison of unpruned counterpart, including ~90.8% decrease of FLOPs, ~92.0% decline of parameter size, running ~2 times faster and comparable detection accuracy as YOLOv3. Experimental results with different pruning ratios consistently verify that proposed SlimYOLOv3 with narrower structure are more efficient, faster and better than YOLOv3, and thus are more suitable for real-time object detection on UAVs. Our codes are made publicly available at https://github.com/PengyiZhang/SlimYOLOv3.

1. Introduction

UAV scene parsing requires object detection on platforms constrained by memory and computing power. The paper addresses this challenge by learning slimmer YOLOv3 detectors through channel pruning, achieving substantial efficiency gains while retaining comparable detection accuracy.

  • UAV applications require scene parsing capabilities including object recognition, localization, and boundary determination.
  • Deep object detectors may contain redundant or non-critical computations that can be removed without significant performance degradation.Channel pruning offers a coarse-grained, implementation-friendly compression approach, followed by fine-tuning.
  • The method imposes L1 regularization on channel scaling factors and removes feature channels with small scaling factors.This produces slim object detectors with fewer informative-channel computations.
  • ~90.8% FLOPs reduction, ~92.0% parameter-size decline, ~2 times faster inference, and comparable detection accuracy are reported for SlimYOLOv3 versus YOLOv3.The evaluation uses the VisDrone2018-Det benchmark and compares against the unpruned counterpart.

2. Related Work

Related work frames model pruning as one form of model compression for resource-limited deployment. Structured channel pruning is favored over unstructured weight pruning because it yields more regular architectures that are easier to implement and accelerate.

  • Deep model compression methods include pruning, knowledge distillation, parameter quantization, and dynamic computation.The paper focuses specifically on model pruning.
  • Incremental pruning evaluates component importance, removes less important components, fine-tunes the model, and evaluates deployment suitability.The procedure is iterative because incremental pruning is preferred to prevent over-pruning.
  • Weight pruning removes small-weight connections but can create irregular architectures that are difficult to store and accelerate.Specialized software or hardware may be needed for practical deployment.
  • Structured pruning produces more regular and tractable network architectures than weight pruning.Structured sparsity can be induced through structured regularization or channel-wise scaling factors.

3. SlimYOLOv3

SlimYOLOv3 searches for compact YOLOv3 channel configurations using sparsity training, channel pruning, and fine-tuning. The design also incorporates SPP modules to enrich multiscale features before pruning.

  • SlimYOLOv3: SlimYOLOv3 applies channel pruning to YOLOv3 to reduce trainable parameters and FLOPs through a compact channel configuration.The procedure follows the iterative process shown in Figure 3.
  • YOLOv3 and SPP: YOLOv3 uses Darknet-53 and three detection scales, while YOLOv3-SPP3 adds spatial pyramid pooling before each detection header.The SPP modules are placed between the fifth and sixth convolutional layers.
  • Sparsity training: Channel-wise sparsity training assigns each channel a scaling factor whose absolute value represents channel importance.Batch-normalization scale factors are used as channel-importance indicators.
  • Sparsity training: L1 regularization on scaling factors encourages sparse channels that can be pruned.The penalty factor α balances the sparsity term and YOLO loss.
  • Channel pruning: Global thresholds based on scaling-factor percentiles control pruning ratios, while local safety thresholds preserve at least 10% of channels per layer.Connected layers use combined pruning masks to maintain shortcut connectivity.
  • Fine-tuning: Fine-tuning retrains the pruned detector to compensate for potentially temporary detection-performance degradation.This step is especially important because fine-grained object detection is sensitive to channel pruning.
  • Iterative pruning: Incremental pruning is preferred because over-pruning can cause catastrophic degradation that the pruned model cannot recover from.

4. Experiments

The experiments evaluate SlimYOLOv3 variants on VisDrone2018-Det against YOLOv3-based baselines, using multiple pruning settings and standard detection, efficiency, and inference metrics.

  • Dataset and baselines: The study evaluates models on VisDrone2018-Det using training and validation splits from the drone-captured benchmark.The dataset contains 7,019 images, with 6,471 for training and 548 for validation, annotated across ten object classes.
  • Dataset and baselines: YOLOv3-tiny serves as a faster but less accurate baseline, while YOLOv3-SPP1 serves as the YOLOv3 baseline.YOLOv3-SPP1 includes one SPP module and is reported as more accurate than original YOLOv3 on COCO.
  • Model variants: YOLOv3-SPP3 adds three SPP modules before the detection headers to seek higher detection accuracy than the baseline models.
  • Pruning settings: SlimYOLOv3-SPP3 variants use 50%, 90%, and 95% global pruning ratios, while a local threshold preserves at least 10% of channels per layer.The pruned models are named SlimYOLOv3-SPP3-50, SlimYOLOv3-SPP3-90, and SlimYOLOv3-SPP3-95.
  • Training procedure: The experimental pipeline includes normal training, 100-epoch sparsity training with three penalty factors, and fine-tuning of the pruned models.Fine-tuning uses pruned YOLOv3-SPP3 weights and adjusts training schedules for different pruning levels.
  • Evaluation: Evaluation covers precision, recall, mAP at 0.5 IOU, F1-score, model volume, parameter size, FLOPs, and inference time in FPS.Experiments use fixed confidence and non-maximum suppression thresholds and no batch processing on an NVIDIA GTX10-series GPU.

5. Results and Discussions

Experiments on VisDrone2018-Det compare SPP variants, sparsity training, and progressively pruned SlimYOLOv3 models. Channel pruning substantially reduces computation, parameters, model volume, and inference time while retaining comparable detection accuracy, although pruning does not fully eliminate real-time bottlenecks and category imbalance remains unaddressed.

  • Effect of SPP modules: At 832×832 input size, YOLOv3-SPP3 outperforms YOLOv3-SPP1 by ~1% in mAP and F1-score, with slightly higher parameters and FLOPs.The result suggests SPP modules help extract multiscale features from high-resolution images.
  • Effect of sparsity training: During sparsity training, smaller scaling factors become more numerous, but α = 0.01 causes overly aggressive decay and underfitting; α = 0.0001 is used for pruning.The scaling-factor histograms and training/validation loss curves are monitored during sparsity training.
  • Effect of channel pruning: 57.1%, 85.9% and 90.8% FLOPs reductions accompany 50%, 90% and 95% channel-pruning ratios, respectively.The comparisons use input size 832×832 and YOLOv3-SPP3 as the baseline.
  • Effect of channel pruning: 67.4%, 87.5% and 92.0% parameter-size reductions and 67.5%, 87.4% and 92.0% model-volume reductions correspond to 50%, 90% and 95% pruning.SlimYOLOv3-SPP3-90 and SlimYOLOv3-SPP3-95 use fewer trainable parameters than YOLOv3-tiny while maintaining comparable detection accuracy to YOLOv3.
  • Effect of channel pruning: 38.8%, 42.6% and 49.5% inference-time reductions correspond to 50%, 90% and 95% pruning, making SlimYOLOv3-SPP3 run ~2 times faster than YOLOv3-SPP3.Inference was measured at 832×832 on an NVIDIA GTX1080ti GPU using Darknet without batch processing.
  • Analysis of detection accuracy: Iterative pruning is more prone to maintaining detection accuracy than aggressive pruning, while channel pruning may encounter a bottleneck in improving real-time performance.SlimYOLOv3-SPP3-95 and YOLOv3-SPP3 detect most objects precisely in the illustrated challenging drone frame without significant difference.

6. Conclusion

SlimYOLOv3 uses channel pruning to produce a narrower detector with fewer parameters and FLOPs than YOLOv3. It maintains comparable detection accuracy, runs faster, and is presented as better suited to real-time UAV applications.

  • Channel-level sparsity is induced with L1 regularization on channel scaling factors, after which less informative channels are pruned.The resulting detector has a narrower structure and fewer trainable parameters than YOLOv3.
  • SlimYOLOv3 achieves comparable detection accuracy to YOLOv3 with significantly fewer FLOPs.
  • SlimYOLOv3 runs faster than original YOLOv3 and is argued to be better for real-time UAV applications.The paper connects lower FLOPs with lower power consumption, which supports UAV endurance.

Inference

The cited inference-related material spans object-detection architectures, segmentation, pruning, quantization, and the VisDrone benchmark. These references situate SlimYOLOv3 among prior approaches to efficient vision and UAV detection.

  • Prior work includes CNN-based object detectors such as R-CNN, YOLO, Faster R-CNN, SSD, and Mask R-CNN.
  • The cited literature also covers efficiency techniques including channel pruning, network slimming, knowledge distillation, quantized CNNs, and structured sparsity.
  • VisDrone-VDT2018 is cited as the drone video detection and tracking challenge associated with the benchmark context.
Loading 1907.11093v1…