What object detection API pipeline settings will likely affect RPN localization loss? - tensorflow

Question

I'm attempting to train a Faster RCNN inception v2 object detector to locate relatively small objects (16x16 with a few at 100x100 pixels) in a set of 512x512 images. I ran with an default configuration and ended up with low accuracy ~0.4 mAP # 0.5 IoU. I then changed some of the pipeline configuration to try to account for the object size but now I'm getting no detections at all (0.00 mAP) and the RPN localization_loss plateaus at 1.0 (the other loss components are significantly lower ~0.05). I have these questions about my pipeline.config updates:
My assumption is that the network is having problems selecting appropriate region proposals. Is that likely, and if so what configuration changes are likely to help?
Did I make some mistake in the changes I made that wrecked the accuracy?
Configuration Changes Attempted:
first_stage_features_stride: Lowered from 16 to 8
grid_anchor_generator: Set anchor size to 64x64 and lowered width/height stride from 16 to 8
train_config: Increased batch_size from 1 to 16
Complete Pipeline Configuration Below:
model {
faster_rcnn {
num_classes: 8
image_resizer {
fixed_shape_resizer {
height: 512
width: 512
}
}
feature_extractor {
type: "faster_rcnn_inception_v2"
first_stage_features_stride: 8
}
first_stage_anchor_generator {
grid_anchor_generator {
height: 64
width: 64
height_stride: 8
width_stride: 8
scales: 0.25
scales: 0.5
scales: 1.0
scales: 2.0
aspect_ratios: 0.5
aspect_ratios: 1.0
aspect_ratios: 2.0
}
}
first_stage_box_predictor_conv_hyperparams {
op: CONV
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
truncated_normal_initializer {
stddev: 0.00999999977648
}
}
}
first_stage_nms_score_threshold: 0.0
first_stage_nms_iou_threshold: 0.699999988079
first_stage_max_proposals: 100
first_stage_localization_loss_weight: 2.0
first_stage_objectness_loss_weight: 1.0
initial_crop_size: 14
maxpool_kernel_size: 2
maxpool_stride: 2
second_stage_box_predictor {
mask_rcnn_box_predictor {
fc_hyperparams {
op: FC
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
variance_scaling_initializer {
factor: 1.0
uniform: true
mode: FAN_AVG
}
}
}
use_dropout: false
dropout_keep_probability: 1.0
}
}
second_stage_post_processing {
batch_non_max_suppression {
score_threshold: 0.300000011921
iou_threshold: 0.600000023842
max_detections_per_class: 100
max_total_detections: 200
}
score_converter: SOFTMAX
}
second_stage_localization_loss_weight: 2.0
second_stage_classification_loss_weight: 1.0
}
}
train_config {
batch_size: 16
data_augmentation_options {
random_horizontal_flip {
}
}
optimizer {
momentum_optimizer {
learning_rate {
manual_step_learning_rate {
initial_learning_rate: 0.000199999994948
schedule {
step: 0
learning_rate: 0.000199999994948
}
schedule {
step: 900000
learning_rate: 1.99999994948e-05
}
schedule {
step: 1200000
learning_rate: 1.99999999495e-06
}
}
}
momentum_optimizer_value: 0.899999976158
}
use_moving_average: false
}
gradient_clipping_by_norm: 10.0
#fine_tune_checkpoint: “XXXXXXXXX”
from_detection_checkpoint: true
#num_steps: 200000
}
train_input_reader {
label_map_path: “XXXXXXXXXX”
tf_record_input_reader {
input_path: “XXXXXXXXXXXX”
}
}
eval_config {
num_examples: 1500
#max_evals: 10
use_moving_averages: false
}
eval_input_reader {
label_map_path: “XXXXXXXXXXXX”
shuffle: false
num_readers: 1
tf_record_input_reader {
input_path: “XXXXXXXXXXXXXba”
}
}