search

shaunyuan22 / CFINet

The official implementation for ICCV'23 paper "Small Object Detection via Coarse-to-fine Proposal Generation and Imitation Learning"
Apache License 2.0
114 stars 5 forks source link

num_base_anchors的值大于1时报错 #24

Closed Renalyn-pan closed 3 months ago

Renalyn-pan commented 3 months ago

当我将anchor_generator中每个锚点生成的锚框设置为3时([0.5, 1.0, 1.5])，在dynamic_assigner中进行assign_wrt_overlaps操作时，在更新最低阈值时会报错inds和potential_low_thr的维度不匹配： IndexError: The shape of the mask [87296] at index 0 does not match the shape of the indexed tensor [261888] at index 0 对应报错代码： for i in range(num_gts): inds = (anchor_argmax_overlaps == i) & \ (anchor_max_overlaps >= gt_pos_thrs[i]) potential_low_thr[inds] = gt_pos_thrs[i] 请问这个应该怎么修改呢？

shaunyuan22 commented 3 months ago

num_base_anchors大于1时，后边的feature adaption没法进行的，因为一个特征点只能对应一对偏移量(delta_x, delta_y)

Renalyn-pan commented 3 months ago

那用于SODA-A遥感中也是只一个特征点设置一个偏移的框是吗，因为普通都是会设置三个长宽比的框

---原始邮件--- 发件人: "Shaun @.> 发送时间: 2024年3月23日(周六) 晚上10:01 收件人: @.>; 抄送: @.**@.>; 主题: Re: [shaunyuan22/CFINet] num_base_anchors的值大于1时报错 (Issue #24)

num_base_anchors大于1时，后边的feature adaption没法进行的，因为一个特征点只能对应一对偏移量(delta_x, delta_y)

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

shaunyuan22 commented 3 months ago

是的，与Cascade RPN保持一致在每个点只放置一个anchor