push custom data

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+custom_data/** filter=lfs diff=lfs merge=lfs -text

README.md

@@ -9,7 +9,7 @@ Jincheng Li*, Chunyu Xie*, Ji Ao, Dawei Leng†, Yuhui Yin (*Equal Contributi
 
 
 ## 🔥 News
-- 🚀 **[2025/07/31]** We have updated the LMM-Det github repository, and now you can test our models!
+- 🚀 **[2025/08/01]** We have updated the LMM-Det github repository, and now you can test our models!
 - 🚀 **[2025/07/24]** We released the paper of [LMM-Det: Make Large Multimodal Models Excel in Object Detection](https://arxiv.org/abs/2507.18300).
 - 🚀 **[2025/06/26]** LMM-Det has been accepted by ICCV'25.
 

custom_data/custom_data.md

@@ -0,0 +1,16 @@
+# Data Curation
+
+In Stage IV, we curate a customized dataset to make LMM-Det excel in object detection while preserving its inherent capabilities like caption generation and VQA.
+
+## Step 1
+
+We generate pesudo labels on the trainset of COCO using [Salience-DETR](https://github.com/xiuqhou/Salience-DETR) (FocalNet-L backone), and re-organize them into a instruction format. Note that the re-organization data consists of ground-truth labels and pesudo labels.
+(In practice, this data is aslo used in Stage III.)
+
+## Step 2
+
+We remove the textcaps data in the LLaVA-665K instruction data.
+
+## Step 3
+
+We concat the the re-organization data and the LLaVA-665K instruction data (without textcaps) as the training data in Stage IV.

custom_data/llava_665k_owlv2_pad_rm_textcaps_w_coco_reorganized_for_stage4.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45e7f79788fd0acf67bdf598ac184c5798907c1f7e58cc83d1d9ea123df67b0b
+size 1306355879