Faster RCNN小目标检测训练异常：RPN定位损失居高不下问询

Let's break down your problem step by step—you're dealing with a classic small-object detection challenge combined with pipeline configuration missteps that caused complete training failure. Here's a targeted analysis and solution:

1. Is Your Hypothesis Valid? (RPN Struggles to Generate Good Proposals) #

Yes, your hypothesis is 100% correct. The stable RPN localization loss of 1.0 (with other losses hovering near 0.05) is a clear red flag that the RPN isn't learning to produce meaningful region proposals for your small targets. Here's why, and how to fix it:

Key Issues & Fixes: #

• Anchor-Target Size Mismatch: Your primary targets are 16x16 pixels, but you set the base anchor size to 64x64. This is drastically misaligned—most of your 16x16 targets will have an IoU < 0.3 with these oversized anchors, meaning the RPN gets almost no positive samples to learn from. Without positive samples, the RPN can't learn to regress accurate boxes, leading to the stuck localization loss and zero detections.
  • Fix: Set your base anchor size to match your dominant target size (16x16), then use scales to cover the full range of your targets (16x16 to 100x100). Example adjustment:

    first_stage_anchor_generator {
      grid_anchor_generator {
        height: 16
        width: 16
        height_stride: 8
        width_stride: 8
        scales: 0.5   # Covers 8x8 tiny targets
        scales: 1.0   # Matches your 16x16 primary targets
        scales: 2.0   # Covers 32x32
        scales: 6.25  # Covers 100x100 (16 * 6.25 = 100)
        aspect_ratios: 0.5
        aspect_ratios: 1.0
        aspect_ratios: 2.0
      }
    }
    

• Feature Stride Adjustment: Lowering first_stage_features_stride from 16 to 8 was a smart move—this gives you a higher-resolution feature map (2x more pixels) which is critical for detecting small objects. Keep this change.
• Positive Sample Matching: The default RPN uses IoU > 0.7 as positive samples. With your old 64x64 anchors, even a 100x100 target would have an IoU of ~0.25 (way below 0.7), so no positive samples were selected. Fixing the anchor size will resolve this.

2. Did Your Configuration Changes Cause the Complete Failure? #

Yes, the anchor size mismatch is the root cause of your mAP dropping to 0.0. The other changes had secondary or no impact, but here's a breakdown:

• Anchor Size (Critical Error): As explained above, 64x64 anchors are completely misaligned with your target sizes, leading to no positive RPN samples. Without valid proposals, the second stage (Fast R-CNN) has nothing to classify, hence zero detections.
• Batch Size Increase: Jumping from batch_size 1 to 16 is a large leap. If your GPU doesn't have enough VRAM, this could cause silent failures (e.g., incomplete weight updates), but this isn't the main issue here. Start with a smaller batch size (4-8) that fits your GPU, and scale your learning rate proportionally (e.g., if you go from batch_size 1 to 8, multiply your initial learning rate by 8).
• Feature Stride Change: This was a correct adjustment for small objects—no issue here, though you may need to retrain the feature extractor's final layers if you're using a pre-trained checkpoint (since the stride change alters the feature map dimensions).

Additional Recommendations: #

• Check Pre-Trained Checkpoint Compatibility: You're using from_detection_checkpoint: true, but the pre-trained Faster R-CNN Inception v2 uses a stride of 16. Changing to stride 8 means the feature extractor's output dimensions are different, so some pre-trained weights may not be compatible. Try training from scratch first (comment out the fine-tune checkpoint lines) to verify your anchor fixes work, then reintroduce fine-tuning with a learning rate tailored to the new stride.
• Adjust RPN NMS Threshold: Your first_stage_nms_iou_threshold is set to 0.7—for small targets, you can lower this slightly (e.g., 0.5) to keep more proposals that might overlap with small objects.
• Monitor Positive Sample Counts: Use TensorBoard to track the number of positive RPN samples per batch. If it's still near zero after fixing anchors, double-check your label format (ensure bounding boxes are in absolute pixel coordinates, normalized correctly, etc.).

内容的提问来源于stack exchange，提问作者J.Doe